1 //===- InstructionCombining.cpp - Combine multiple instructions -----------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // InstructionCombining - Combine instructions to form fewer, simple
11 // instructions. This pass does not modify the CFG This pass is where algebraic
12 // simplification happens.
14 // This pass combines things like:
20 // This is a simple worklist driven algorithm.
22 // This pass guarantees that the following canonicalizations are performed on
24 // 1. If a binary operator has a constant operand, it is moved to the RHS
25 // 2. Bitwise operators with constant operands are always grouped so that
26 // shifts are performed first, then or's, then and's, then xor's.
27 // 3. Compare instructions are converted from <,>,<=,>= to ==,!= if possible
28 // 4. All cmp instructions on boolean values are replaced with logical ops
29 // 5. add X, X is represented as (X*2) => (X << 1)
30 // 6. Multiplies with a power-of-two constant argument are transformed into
34 //===----------------------------------------------------------------------===//
36 #define DEBUG_TYPE "instcombine"
37 #include "llvm/Transforms/Scalar.h"
38 #include "llvm/IntrinsicInst.h"
39 #include "llvm/Pass.h"
40 #include "llvm/DerivedTypes.h"
41 #include "llvm/GlobalVariable.h"
42 #include "llvm/ParameterAttributes.h"
43 #include "llvm/Analysis/ConstantFolding.h"
44 #include "llvm/Target/TargetData.h"
45 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
46 #include "llvm/Transforms/Utils/Local.h"
47 #include "llvm/Support/CallSite.h"
48 #include "llvm/Support/ConstantRange.h"
49 #include "llvm/Support/Debug.h"
50 #include "llvm/Support/GetElementPtrTypeIterator.h"
51 #include "llvm/Support/InstVisitor.h"
52 #include "llvm/Support/MathExtras.h"
53 #include "llvm/Support/PatternMatch.h"
54 #include "llvm/Support/Compiler.h"
55 #include "llvm/ADT/DenseMap.h"
56 #include "llvm/ADT/SmallVector.h"
57 #include "llvm/ADT/SmallPtrSet.h"
58 #include "llvm/ADT/Statistic.h"
59 #include "llvm/ADT/STLExtras.h"
63 using namespace llvm::PatternMatch;
65 STATISTIC(NumCombined , "Number of insts combined");
66 STATISTIC(NumConstProp, "Number of constant folds");
67 STATISTIC(NumDeadInst , "Number of dead inst eliminated");
68 STATISTIC(NumDeadStore, "Number of dead stores eliminated");
69 STATISTIC(NumSunkInst , "Number of instructions sunk");
72 class VISIBILITY_HIDDEN InstCombiner
73 : public FunctionPass,
74 public InstVisitor<InstCombiner, Instruction*> {
75 // Worklist of all of the instructions that need to be simplified.
76 std::vector<Instruction*> Worklist;
77 DenseMap<Instruction*, unsigned> WorklistMap;
79 bool MustPreserveLCSSA;
81 static char ID; // Pass identification, replacement for typeid
82 InstCombiner() : FunctionPass((intptr_t)&ID) {}
84 /// AddToWorkList - Add the specified instruction to the worklist if it
85 /// isn't already in it.
86 void AddToWorkList(Instruction *I) {
87 if (WorklistMap.insert(std::make_pair(I, Worklist.size())))
88 Worklist.push_back(I);
91 // RemoveFromWorkList - remove I from the worklist if it exists.
92 void RemoveFromWorkList(Instruction *I) {
93 DenseMap<Instruction*, unsigned>::iterator It = WorklistMap.find(I);
94 if (It == WorklistMap.end()) return; // Not in worklist.
96 // Don't bother moving everything down, just null out the slot.
97 Worklist[It->second] = 0;
99 WorklistMap.erase(It);
102 Instruction *RemoveOneFromWorkList() {
103 Instruction *I = Worklist.back();
105 WorklistMap.erase(I);
110 /// AddUsersToWorkList - When an instruction is simplified, add all users of
111 /// the instruction to the work lists because they might get more simplified
114 void AddUsersToWorkList(Value &I) {
115 for (Value::use_iterator UI = I.use_begin(), UE = I.use_end();
117 AddToWorkList(cast<Instruction>(*UI));
120 /// AddUsesToWorkList - When an instruction is simplified, add operands to
121 /// the work lists because they might get more simplified now.
123 void AddUsesToWorkList(Instruction &I) {
124 for (unsigned i = 0, e = I.getNumOperands(); i != e; ++i)
125 if (Instruction *Op = dyn_cast<Instruction>(I.getOperand(i)))
129 /// AddSoonDeadInstToWorklist - The specified instruction is about to become
130 /// dead. Add all of its operands to the worklist, turning them into
131 /// undef's to reduce the number of uses of those instructions.
133 /// Return the specified operand before it is turned into an undef.
135 Value *AddSoonDeadInstToWorklist(Instruction &I, unsigned op) {
136 Value *R = I.getOperand(op);
138 for (unsigned i = 0, e = I.getNumOperands(); i != e; ++i)
139 if (Instruction *Op = dyn_cast<Instruction>(I.getOperand(i))) {
141 // Set the operand to undef to drop the use.
142 I.setOperand(i, UndefValue::get(Op->getType()));
149 virtual bool runOnFunction(Function &F);
151 bool DoOneIteration(Function &F, unsigned ItNum);
153 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
154 AU.addRequired<TargetData>();
155 AU.addPreservedID(LCSSAID);
156 AU.setPreservesCFG();
159 TargetData &getTargetData() const { return *TD; }
161 // Visitation implementation - Implement instruction combining for different
162 // instruction types. The semantics are as follows:
164 // null - No change was made
165 // I - Change was made, I is still valid, I may be dead though
166 // otherwise - Change was made, replace I with returned instruction
168 Instruction *visitAdd(BinaryOperator &I);
169 Instruction *visitSub(BinaryOperator &I);
170 Instruction *visitMul(BinaryOperator &I);
171 Instruction *visitURem(BinaryOperator &I);
172 Instruction *visitSRem(BinaryOperator &I);
173 Instruction *visitFRem(BinaryOperator &I);
174 Instruction *commonRemTransforms(BinaryOperator &I);
175 Instruction *commonIRemTransforms(BinaryOperator &I);
176 Instruction *commonDivTransforms(BinaryOperator &I);
177 Instruction *commonIDivTransforms(BinaryOperator &I);
178 Instruction *visitUDiv(BinaryOperator &I);
179 Instruction *visitSDiv(BinaryOperator &I);
180 Instruction *visitFDiv(BinaryOperator &I);
181 Instruction *visitAnd(BinaryOperator &I);
182 Instruction *visitOr (BinaryOperator &I);
183 Instruction *visitXor(BinaryOperator &I);
184 Instruction *visitShl(BinaryOperator &I);
185 Instruction *visitAShr(BinaryOperator &I);
186 Instruction *visitLShr(BinaryOperator &I);
187 Instruction *commonShiftTransforms(BinaryOperator &I);
188 Instruction *visitFCmpInst(FCmpInst &I);
189 Instruction *visitICmpInst(ICmpInst &I);
190 Instruction *visitICmpInstWithCastAndCast(ICmpInst &ICI);
191 Instruction *visitICmpInstWithInstAndIntCst(ICmpInst &ICI,
194 Instruction *FoldICmpDivCst(ICmpInst &ICI, BinaryOperator *DivI,
195 ConstantInt *DivRHS);
197 Instruction *FoldGEPICmp(User *GEPLHS, Value *RHS,
198 ICmpInst::Predicate Cond, Instruction &I);
199 Instruction *FoldShiftByConstant(Value *Op0, ConstantInt *Op1,
201 Instruction *commonCastTransforms(CastInst &CI);
202 Instruction *commonIntCastTransforms(CastInst &CI);
203 Instruction *commonPointerCastTransforms(CastInst &CI);
204 Instruction *visitTrunc(TruncInst &CI);
205 Instruction *visitZExt(ZExtInst &CI);
206 Instruction *visitSExt(SExtInst &CI);
207 Instruction *visitFPTrunc(FPTruncInst &CI);
208 Instruction *visitFPExt(CastInst &CI);
209 Instruction *visitFPToUI(CastInst &CI);
210 Instruction *visitFPToSI(CastInst &CI);
211 Instruction *visitUIToFP(CastInst &CI);
212 Instruction *visitSIToFP(CastInst &CI);
213 Instruction *visitPtrToInt(CastInst &CI);
214 Instruction *visitIntToPtr(IntToPtrInst &CI);
215 Instruction *visitBitCast(BitCastInst &CI);
216 Instruction *FoldSelectOpOp(SelectInst &SI, Instruction *TI,
218 Instruction *visitSelectInst(SelectInst &CI);
219 Instruction *visitCallInst(CallInst &CI);
220 Instruction *visitInvokeInst(InvokeInst &II);
221 Instruction *visitPHINode(PHINode &PN);
222 Instruction *visitGetElementPtrInst(GetElementPtrInst &GEP);
223 Instruction *visitAllocationInst(AllocationInst &AI);
224 Instruction *visitFreeInst(FreeInst &FI);
225 Instruction *visitLoadInst(LoadInst &LI);
226 Instruction *visitStoreInst(StoreInst &SI);
227 Instruction *visitBranchInst(BranchInst &BI);
228 Instruction *visitSwitchInst(SwitchInst &SI);
229 Instruction *visitInsertElementInst(InsertElementInst &IE);
230 Instruction *visitExtractElementInst(ExtractElementInst &EI);
231 Instruction *visitShuffleVectorInst(ShuffleVectorInst &SVI);
233 // visitInstruction - Specify what to return for unhandled instructions...
234 Instruction *visitInstruction(Instruction &I) { return 0; }
237 Instruction *visitCallSite(CallSite CS);
238 bool transformConstExprCastCall(CallSite CS);
239 Instruction *transformCallThroughTrampoline(CallSite CS);
242 // InsertNewInstBefore - insert an instruction New before instruction Old
243 // in the program. Add the new instruction to the worklist.
245 Instruction *InsertNewInstBefore(Instruction *New, Instruction &Old) {
246 assert(New && New->getParent() == 0 &&
247 "New instruction already inserted into a basic block!");
248 BasicBlock *BB = Old.getParent();
249 BB->getInstList().insert(&Old, New); // Insert inst
254 /// InsertCastBefore - Insert a cast of V to TY before the instruction POS.
255 /// This also adds the cast to the worklist. Finally, this returns the
257 Value *InsertCastBefore(Instruction::CastOps opc, Value *V, const Type *Ty,
259 if (V->getType() == Ty) return V;
261 if (Constant *CV = dyn_cast<Constant>(V))
262 return ConstantExpr::getCast(opc, CV, Ty);
264 Instruction *C = CastInst::create(opc, V, Ty, V->getName(), &Pos);
269 Value *InsertBitCastBefore(Value *V, const Type *Ty, Instruction &Pos) {
270 return InsertCastBefore(Instruction::BitCast, V, Ty, Pos);
274 // ReplaceInstUsesWith - This method is to be used when an instruction is
275 // found to be dead, replacable with another preexisting expression. Here
276 // we add all uses of I to the worklist, replace all uses of I with the new
277 // value, then return I, so that the inst combiner will know that I was
280 Instruction *ReplaceInstUsesWith(Instruction &I, Value *V) {
281 AddUsersToWorkList(I); // Add all modified instrs to worklist
283 I.replaceAllUsesWith(V);
286 // If we are replacing the instruction with itself, this must be in a
287 // segment of unreachable code, so just clobber the instruction.
288 I.replaceAllUsesWith(UndefValue::get(I.getType()));
293 // UpdateValueUsesWith - This method is to be used when an value is
294 // found to be replacable with another preexisting expression or was
295 // updated. Here we add all uses of I to the worklist, replace all uses of
296 // I with the new value (unless the instruction was just updated), then
297 // return true, so that the inst combiner will know that I was modified.
299 bool UpdateValueUsesWith(Value *Old, Value *New) {
300 AddUsersToWorkList(*Old); // Add all modified instrs to worklist
302 Old->replaceAllUsesWith(New);
303 if (Instruction *I = dyn_cast<Instruction>(Old))
305 if (Instruction *I = dyn_cast<Instruction>(New))
310 // EraseInstFromFunction - When dealing with an instruction that has side
311 // effects or produces a void value, we can't rely on DCE to delete the
312 // instruction. Instead, visit methods should return the value returned by
314 Instruction *EraseInstFromFunction(Instruction &I) {
315 assert(I.use_empty() && "Cannot erase instruction that is used!");
316 AddUsesToWorkList(I);
317 RemoveFromWorkList(&I);
319 return 0; // Don't do anything with FI
323 /// InsertOperandCastBefore - This inserts a cast of V to DestTy before the
324 /// InsertBefore instruction. This is specialized a bit to avoid inserting
325 /// casts that are known to not do anything...
327 Value *InsertOperandCastBefore(Instruction::CastOps opcode,
328 Value *V, const Type *DestTy,
329 Instruction *InsertBefore);
331 /// SimplifyCommutative - This performs a few simplifications for
332 /// commutative operators.
333 bool SimplifyCommutative(BinaryOperator &I);
335 /// SimplifyCompare - This reorders the operands of a CmpInst to get them in
336 /// most-complex to least-complex order.
337 bool SimplifyCompare(CmpInst &I);
339 /// SimplifyDemandedBits - Attempts to replace V with a simpler value based
340 /// on the demanded bits.
341 bool SimplifyDemandedBits(Value *V, APInt DemandedMask,
342 APInt& KnownZero, APInt& KnownOne,
345 Value *SimplifyDemandedVectorElts(Value *V, uint64_t DemandedElts,
346 uint64_t &UndefElts, unsigned Depth = 0);
348 // FoldOpIntoPhi - Given a binary operator or cast instruction which has a
349 // PHI node as operand #0, see if we can fold the instruction into the PHI
350 // (which is only possible if all operands to the PHI are constants).
351 Instruction *FoldOpIntoPhi(Instruction &I);
353 // FoldPHIArgOpIntoPHI - If all operands to a PHI node are the same "unary"
354 // operator and they all are only used by the PHI, PHI together their
355 // inputs, and do the operation once, to the result of the PHI.
356 Instruction *FoldPHIArgOpIntoPHI(PHINode &PN);
357 Instruction *FoldPHIArgBinOpIntoPHI(PHINode &PN);
360 Instruction *OptAndOp(Instruction *Op, ConstantInt *OpRHS,
361 ConstantInt *AndRHS, BinaryOperator &TheAnd);
363 Value *FoldLogicalPlusAnd(Value *LHS, Value *RHS, ConstantInt *Mask,
364 bool isSub, Instruction &I);
365 Instruction *InsertRangeTest(Value *V, Constant *Lo, Constant *Hi,
366 bool isSigned, bool Inside, Instruction &IB);
367 Instruction *PromoteCastOfAllocation(BitCastInst &CI, AllocationInst &AI);
368 Instruction *MatchBSwap(BinaryOperator &I);
369 bool SimplifyStoreAtEndOfBlock(StoreInst &SI);
370 Instruction *SimplifyMemTransfer(MemIntrinsic *MI);
373 Value *EvaluateInDifferentType(Value *V, const Type *Ty, bool isSigned);
376 char InstCombiner::ID = 0;
377 RegisterPass<InstCombiner> X("instcombine", "Combine redundant instructions");
380 // getComplexity: Assign a complexity or rank value to LLVM Values...
381 // 0 -> undef, 1 -> Const, 2 -> Other, 3 -> Arg, 3 -> Unary, 4 -> OtherInst
382 static unsigned getComplexity(Value *V) {
383 if (isa<Instruction>(V)) {
384 if (BinaryOperator::isNeg(V) || BinaryOperator::isNot(V))
388 if (isa<Argument>(V)) return 3;
389 return isa<Constant>(V) ? (isa<UndefValue>(V) ? 0 : 1) : 2;
392 // isOnlyUse - Return true if this instruction will be deleted if we stop using
394 static bool isOnlyUse(Value *V) {
395 return V->hasOneUse() || isa<Constant>(V);
398 // getPromotedType - Return the specified type promoted as it would be to pass
399 // though a va_arg area...
400 static const Type *getPromotedType(const Type *Ty) {
401 if (const IntegerType* ITy = dyn_cast<IntegerType>(Ty)) {
402 if (ITy->getBitWidth() < 32)
403 return Type::Int32Ty;
408 /// getBitCastOperand - If the specified operand is a CastInst or a constant
409 /// expression bitcast, return the operand value, otherwise return null.
410 static Value *getBitCastOperand(Value *V) {
411 if (BitCastInst *I = dyn_cast<BitCastInst>(V))
412 return I->getOperand(0);
413 else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
414 if (CE->getOpcode() == Instruction::BitCast)
415 return CE->getOperand(0);
419 /// This function is a wrapper around CastInst::isEliminableCastPair. It
420 /// simply extracts arguments and returns what that function returns.
421 static Instruction::CastOps
422 isEliminableCastPair(
423 const CastInst *CI, ///< The first cast instruction
424 unsigned opcode, ///< The opcode of the second cast instruction
425 const Type *DstTy, ///< The target type for the second cast instruction
426 TargetData *TD ///< The target data for pointer size
429 const Type *SrcTy = CI->getOperand(0)->getType(); // A from above
430 const Type *MidTy = CI->getType(); // B from above
432 // Get the opcodes of the two Cast instructions
433 Instruction::CastOps firstOp = Instruction::CastOps(CI->getOpcode());
434 Instruction::CastOps secondOp = Instruction::CastOps(opcode);
436 return Instruction::CastOps(
437 CastInst::isEliminableCastPair(firstOp, secondOp, SrcTy, MidTy,
438 DstTy, TD->getIntPtrType()));
441 /// ValueRequiresCast - Return true if the cast from "V to Ty" actually results
442 /// in any code being generated. It does not require codegen if V is simple
443 /// enough or if the cast can be folded into other casts.
444 static bool ValueRequiresCast(Instruction::CastOps opcode, const Value *V,
445 const Type *Ty, TargetData *TD) {
446 if (V->getType() == Ty || isa<Constant>(V)) return false;
448 // If this is another cast that can be eliminated, it isn't codegen either.
449 if (const CastInst *CI = dyn_cast<CastInst>(V))
450 if (isEliminableCastPair(CI, opcode, Ty, TD))
455 /// InsertOperandCastBefore - This inserts a cast of V to DestTy before the
456 /// InsertBefore instruction. This is specialized a bit to avoid inserting
457 /// casts that are known to not do anything...
459 Value *InstCombiner::InsertOperandCastBefore(Instruction::CastOps opcode,
460 Value *V, const Type *DestTy,
461 Instruction *InsertBefore) {
462 if (V->getType() == DestTy) return V;
463 if (Constant *C = dyn_cast<Constant>(V))
464 return ConstantExpr::getCast(opcode, C, DestTy);
466 return InsertCastBefore(opcode, V, DestTy, *InsertBefore);
469 // SimplifyCommutative - This performs a few simplifications for commutative
472 // 1. Order operands such that they are listed from right (least complex) to
473 // left (most complex). This puts constants before unary operators before
476 // 2. Transform: (op (op V, C1), C2) ==> (op V, (op C1, C2))
477 // 3. Transform: (op (op V1, C1), (op V2, C2)) ==> (op (op V1, V2), (op C1,C2))
479 bool InstCombiner::SimplifyCommutative(BinaryOperator &I) {
480 bool Changed = false;
481 if (getComplexity(I.getOperand(0)) < getComplexity(I.getOperand(1)))
482 Changed = !I.swapOperands();
484 if (!I.isAssociative()) return Changed;
485 Instruction::BinaryOps Opcode = I.getOpcode();
486 if (BinaryOperator *Op = dyn_cast<BinaryOperator>(I.getOperand(0)))
487 if (Op->getOpcode() == Opcode && isa<Constant>(Op->getOperand(1))) {
488 if (isa<Constant>(I.getOperand(1))) {
489 Constant *Folded = ConstantExpr::get(I.getOpcode(),
490 cast<Constant>(I.getOperand(1)),
491 cast<Constant>(Op->getOperand(1)));
492 I.setOperand(0, Op->getOperand(0));
493 I.setOperand(1, Folded);
495 } else if (BinaryOperator *Op1=dyn_cast<BinaryOperator>(I.getOperand(1)))
496 if (Op1->getOpcode() == Opcode && isa<Constant>(Op1->getOperand(1)) &&
497 isOnlyUse(Op) && isOnlyUse(Op1)) {
498 Constant *C1 = cast<Constant>(Op->getOperand(1));
499 Constant *C2 = cast<Constant>(Op1->getOperand(1));
501 // Fold (op (op V1, C1), (op V2, C2)) ==> (op (op V1, V2), (op C1,C2))
502 Constant *Folded = ConstantExpr::get(I.getOpcode(), C1, C2);
503 Instruction *New = BinaryOperator::create(Opcode, Op->getOperand(0),
507 I.setOperand(0, New);
508 I.setOperand(1, Folded);
515 /// SimplifyCompare - For a CmpInst this function just orders the operands
516 /// so that theyare listed from right (least complex) to left (most complex).
517 /// This puts constants before unary operators before binary operators.
518 bool InstCombiner::SimplifyCompare(CmpInst &I) {
519 if (getComplexity(I.getOperand(0)) >= getComplexity(I.getOperand(1)))
522 // Compare instructions are not associative so there's nothing else we can do.
526 // dyn_castNegVal - Given a 'sub' instruction, return the RHS of the instruction
527 // if the LHS is a constant zero (which is the 'negate' form).
529 static inline Value *dyn_castNegVal(Value *V) {
530 if (BinaryOperator::isNeg(V))
531 return BinaryOperator::getNegArgument(V);
533 // Constants can be considered to be negated values if they can be folded.
534 if (ConstantInt *C = dyn_cast<ConstantInt>(V))
535 return ConstantExpr::getNeg(C);
539 static inline Value *dyn_castNotVal(Value *V) {
540 if (BinaryOperator::isNot(V))
541 return BinaryOperator::getNotArgument(V);
543 // Constants can be considered to be not'ed values...
544 if (ConstantInt *C = dyn_cast<ConstantInt>(V))
545 return ConstantInt::get(~C->getValue());
549 // dyn_castFoldableMul - If this value is a multiply that can be folded into
550 // other computations (because it has a constant operand), return the
551 // non-constant operand of the multiply, and set CST to point to the multiplier.
552 // Otherwise, return null.
554 static inline Value *dyn_castFoldableMul(Value *V, ConstantInt *&CST) {
555 if (V->hasOneUse() && V->getType()->isInteger())
556 if (Instruction *I = dyn_cast<Instruction>(V)) {
557 if (I->getOpcode() == Instruction::Mul)
558 if ((CST = dyn_cast<ConstantInt>(I->getOperand(1))))
559 return I->getOperand(0);
560 if (I->getOpcode() == Instruction::Shl)
561 if ((CST = dyn_cast<ConstantInt>(I->getOperand(1)))) {
562 // The multiplier is really 1 << CST.
563 uint32_t BitWidth = cast<IntegerType>(V->getType())->getBitWidth();
564 uint32_t CSTVal = CST->getLimitedValue(BitWidth);
565 CST = ConstantInt::get(APInt(BitWidth, 1).shl(CSTVal));
566 return I->getOperand(0);
572 /// dyn_castGetElementPtr - If this is a getelementptr instruction or constant
573 /// expression, return it.
574 static User *dyn_castGetElementPtr(Value *V) {
575 if (isa<GetElementPtrInst>(V)) return cast<User>(V);
576 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
577 if (CE->getOpcode() == Instruction::GetElementPtr)
578 return cast<User>(V);
582 /// AddOne - Add one to a ConstantInt
583 static ConstantInt *AddOne(ConstantInt *C) {
584 APInt Val(C->getValue());
585 return ConstantInt::get(++Val);
587 /// SubOne - Subtract one from a ConstantInt
588 static ConstantInt *SubOne(ConstantInt *C) {
589 APInt Val(C->getValue());
590 return ConstantInt::get(--Val);
592 /// Add - Add two ConstantInts together
593 static ConstantInt *Add(ConstantInt *C1, ConstantInt *C2) {
594 return ConstantInt::get(C1->getValue() + C2->getValue());
596 /// And - Bitwise AND two ConstantInts together
597 static ConstantInt *And(ConstantInt *C1, ConstantInt *C2) {
598 return ConstantInt::get(C1->getValue() & C2->getValue());
600 /// Subtract - Subtract one ConstantInt from another
601 static ConstantInt *Subtract(ConstantInt *C1, ConstantInt *C2) {
602 return ConstantInt::get(C1->getValue() - C2->getValue());
604 /// Multiply - Multiply two ConstantInts together
605 static ConstantInt *Multiply(ConstantInt *C1, ConstantInt *C2) {
606 return ConstantInt::get(C1->getValue() * C2->getValue());
609 /// ComputeMaskedBits - Determine which of the bits specified in Mask are
610 /// known to be either zero or one and return them in the KnownZero/KnownOne
611 /// bit sets. This code only analyzes bits in Mask, in order to short-circuit
613 /// NOTE: we cannot consider 'undef' to be "IsZero" here. The problem is that
614 /// we cannot optimize based on the assumption that it is zero without changing
615 /// it to be an explicit zero. If we don't change it to zero, other code could
616 /// optimized based on the contradictory assumption that it is non-zero.
617 /// Because instcombine aggressively folds operations with undef args anyway,
618 /// this won't lose us code quality.
619 static void ComputeMaskedBits(Value *V, const APInt &Mask, APInt& KnownZero,
620 APInt& KnownOne, unsigned Depth = 0) {
621 assert(V && "No Value?");
622 assert(Depth <= 6 && "Limit Search Depth");
623 uint32_t BitWidth = Mask.getBitWidth();
624 assert(cast<IntegerType>(V->getType())->getBitWidth() == BitWidth &&
625 KnownZero.getBitWidth() == BitWidth &&
626 KnownOne.getBitWidth() == BitWidth &&
627 "V, Mask, KnownOne and KnownZero should have same BitWidth");
628 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
629 // We know all of the bits for a constant!
630 KnownOne = CI->getValue() & Mask;
631 KnownZero = ~KnownOne & Mask;
635 if (Depth == 6 || Mask == 0)
636 return; // Limit search depth.
638 Instruction *I = dyn_cast<Instruction>(V);
641 KnownZero.clear(); KnownOne.clear(); // Don't know anything.
642 APInt KnownZero2(KnownZero), KnownOne2(KnownOne);
644 switch (I->getOpcode()) {
645 case Instruction::And: {
646 // If either the LHS or the RHS are Zero, the result is zero.
647 ComputeMaskedBits(I->getOperand(1), Mask, KnownZero, KnownOne, Depth+1);
648 APInt Mask2(Mask & ~KnownZero);
649 ComputeMaskedBits(I->getOperand(0), Mask2, KnownZero2, KnownOne2, Depth+1);
650 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
651 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
653 // Output known-1 bits are only known if set in both the LHS & RHS.
654 KnownOne &= KnownOne2;
655 // Output known-0 are known to be clear if zero in either the LHS | RHS.
656 KnownZero |= KnownZero2;
659 case Instruction::Or: {
660 ComputeMaskedBits(I->getOperand(1), Mask, KnownZero, KnownOne, Depth+1);
661 APInt Mask2(Mask & ~KnownOne);
662 ComputeMaskedBits(I->getOperand(0), Mask2, KnownZero2, KnownOne2, Depth+1);
663 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
664 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
666 // Output known-0 bits are only known if clear in both the LHS & RHS.
667 KnownZero &= KnownZero2;
668 // Output known-1 are known to be set if set in either the LHS | RHS.
669 KnownOne |= KnownOne2;
672 case Instruction::Xor: {
673 ComputeMaskedBits(I->getOperand(1), Mask, KnownZero, KnownOne, Depth+1);
674 ComputeMaskedBits(I->getOperand(0), Mask, KnownZero2, KnownOne2, Depth+1);
675 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
676 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
678 // Output known-0 bits are known if clear or set in both the LHS & RHS.
679 APInt KnownZeroOut = (KnownZero & KnownZero2) | (KnownOne & KnownOne2);
680 // Output known-1 are known to be set if set in only one of the LHS, RHS.
681 KnownOne = (KnownZero & KnownOne2) | (KnownOne & KnownZero2);
682 KnownZero = KnownZeroOut;
685 case Instruction::Select:
686 ComputeMaskedBits(I->getOperand(2), Mask, KnownZero, KnownOne, Depth+1);
687 ComputeMaskedBits(I->getOperand(1), Mask, KnownZero2, KnownOne2, Depth+1);
688 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
689 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
691 // Only known if known in both the LHS and RHS.
692 KnownOne &= KnownOne2;
693 KnownZero &= KnownZero2;
695 case Instruction::FPTrunc:
696 case Instruction::FPExt:
697 case Instruction::FPToUI:
698 case Instruction::FPToSI:
699 case Instruction::SIToFP:
700 case Instruction::PtrToInt:
701 case Instruction::UIToFP:
702 case Instruction::IntToPtr:
703 return; // Can't work with floating point or pointers
704 case Instruction::Trunc: {
705 // All these have integer operands
706 uint32_t SrcBitWidth =
707 cast<IntegerType>(I->getOperand(0)->getType())->getBitWidth();
709 MaskIn.zext(SrcBitWidth);
710 KnownZero.zext(SrcBitWidth);
711 KnownOne.zext(SrcBitWidth);
712 ComputeMaskedBits(I->getOperand(0), MaskIn, KnownZero, KnownOne, Depth+1);
713 KnownZero.trunc(BitWidth);
714 KnownOne.trunc(BitWidth);
717 case Instruction::BitCast: {
718 const Type *SrcTy = I->getOperand(0)->getType();
719 if (SrcTy->isInteger()) {
720 ComputeMaskedBits(I->getOperand(0), Mask, KnownZero, KnownOne, Depth+1);
725 case Instruction::ZExt: {
726 // Compute the bits in the result that are not present in the input.
727 const IntegerType *SrcTy = cast<IntegerType>(I->getOperand(0)->getType());
728 uint32_t SrcBitWidth = SrcTy->getBitWidth();
731 MaskIn.trunc(SrcBitWidth);
732 KnownZero.trunc(SrcBitWidth);
733 KnownOne.trunc(SrcBitWidth);
734 ComputeMaskedBits(I->getOperand(0), MaskIn, KnownZero, KnownOne, Depth+1);
735 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
736 // The top bits are known to be zero.
737 KnownZero.zext(BitWidth);
738 KnownOne.zext(BitWidth);
739 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth);
742 case Instruction::SExt: {
743 // Compute the bits in the result that are not present in the input.
744 const IntegerType *SrcTy = cast<IntegerType>(I->getOperand(0)->getType());
745 uint32_t SrcBitWidth = SrcTy->getBitWidth();
748 MaskIn.trunc(SrcBitWidth);
749 KnownZero.trunc(SrcBitWidth);
750 KnownOne.trunc(SrcBitWidth);
751 ComputeMaskedBits(I->getOperand(0), MaskIn, KnownZero, KnownOne, Depth+1);
752 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
753 KnownZero.zext(BitWidth);
754 KnownOne.zext(BitWidth);
756 // If the sign bit of the input is known set or clear, then we know the
757 // top bits of the result.
758 if (KnownZero[SrcBitWidth-1]) // Input sign bit known zero
759 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth);
760 else if (KnownOne[SrcBitWidth-1]) // Input sign bit known set
761 KnownOne |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth);
764 case Instruction::Shl:
765 // (shl X, C1) & C2 == 0 iff (X & C2 >>u C1) == 0
766 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
767 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth);
768 APInt Mask2(Mask.lshr(ShiftAmt));
769 ComputeMaskedBits(I->getOperand(0), Mask2, KnownZero, KnownOne, Depth+1);
770 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
771 KnownZero <<= ShiftAmt;
772 KnownOne <<= ShiftAmt;
773 KnownZero |= APInt::getLowBitsSet(BitWidth, ShiftAmt); // low bits known 0
777 case Instruction::LShr:
778 // (ushr X, C1) & C2 == 0 iff (-1 >> C1) & C2 == 0
779 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
780 // Compute the new bits that are at the top now.
781 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth);
783 // Unsigned shift right.
784 APInt Mask2(Mask.shl(ShiftAmt));
785 ComputeMaskedBits(I->getOperand(0), Mask2, KnownZero,KnownOne,Depth+1);
786 assert((KnownZero & KnownOne) == 0&&"Bits known to be one AND zero?");
787 KnownZero = APIntOps::lshr(KnownZero, ShiftAmt);
788 KnownOne = APIntOps::lshr(KnownOne, ShiftAmt);
789 // high bits known zero.
790 KnownZero |= APInt::getHighBitsSet(BitWidth, ShiftAmt);
794 case Instruction::AShr:
795 // (ashr X, C1) & C2 == 0 iff (-1 >> C1) & C2 == 0
796 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
797 // Compute the new bits that are at the top now.
798 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth);
800 // Signed shift right.
801 APInt Mask2(Mask.shl(ShiftAmt));
802 ComputeMaskedBits(I->getOperand(0), Mask2, KnownZero,KnownOne,Depth+1);
803 assert((KnownZero & KnownOne) == 0&&"Bits known to be one AND zero?");
804 KnownZero = APIntOps::lshr(KnownZero, ShiftAmt);
805 KnownOne = APIntOps::lshr(KnownOne, ShiftAmt);
807 APInt HighBits(APInt::getHighBitsSet(BitWidth, ShiftAmt));
808 if (KnownZero[BitWidth-ShiftAmt-1]) // New bits are known zero.
809 KnownZero |= HighBits;
810 else if (KnownOne[BitWidth-ShiftAmt-1]) // New bits are known one.
811 KnownOne |= HighBits;
818 /// MaskedValueIsZero - Return true if 'V & Mask' is known to be zero. We use
819 /// this predicate to simplify operations downstream. Mask is known to be zero
820 /// for bits that V cannot have.
821 static bool MaskedValueIsZero(Value *V, const APInt& Mask, unsigned Depth = 0) {
822 APInt KnownZero(Mask.getBitWidth(), 0), KnownOne(Mask.getBitWidth(), 0);
823 ComputeMaskedBits(V, Mask, KnownZero, KnownOne, Depth);
824 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
825 return (KnownZero & Mask) == Mask;
828 /// ShrinkDemandedConstant - Check to see if the specified operand of the
829 /// specified instruction is a constant integer. If so, check to see if there
830 /// are any bits set in the constant that are not demanded. If so, shrink the
831 /// constant and return true.
832 static bool ShrinkDemandedConstant(Instruction *I, unsigned OpNo,
834 assert(I && "No instruction?");
835 assert(OpNo < I->getNumOperands() && "Operand index too large");
837 // If the operand is not a constant integer, nothing to do.
838 ConstantInt *OpC = dyn_cast<ConstantInt>(I->getOperand(OpNo));
839 if (!OpC) return false;
841 // If there are no bits set that aren't demanded, nothing to do.
842 Demanded.zextOrTrunc(OpC->getValue().getBitWidth());
843 if ((~Demanded & OpC->getValue()) == 0)
846 // This instruction is producing bits that are not demanded. Shrink the RHS.
847 Demanded &= OpC->getValue();
848 I->setOperand(OpNo, ConstantInt::get(Demanded));
852 // ComputeSignedMinMaxValuesFromKnownBits - Given a signed integer type and a
853 // set of known zero and one bits, compute the maximum and minimum values that
854 // could have the specified known zero and known one bits, returning them in
856 static void ComputeSignedMinMaxValuesFromKnownBits(const Type *Ty,
857 const APInt& KnownZero,
858 const APInt& KnownOne,
859 APInt& Min, APInt& Max) {
860 uint32_t BitWidth = cast<IntegerType>(Ty)->getBitWidth();
861 assert(KnownZero.getBitWidth() == BitWidth &&
862 KnownOne.getBitWidth() == BitWidth &&
863 Min.getBitWidth() == BitWidth && Max.getBitWidth() == BitWidth &&
864 "Ty, KnownZero, KnownOne and Min, Max must have equal bitwidth.");
865 APInt UnknownBits = ~(KnownZero|KnownOne);
867 // The minimum value is when all unknown bits are zeros, EXCEPT for the sign
868 // bit if it is unknown.
870 Max = KnownOne|UnknownBits;
872 if (UnknownBits[BitWidth-1]) { // Sign bit is unknown
874 Max.clear(BitWidth-1);
878 // ComputeUnsignedMinMaxValuesFromKnownBits - Given an unsigned integer type and
879 // a set of known zero and one bits, compute the maximum and minimum values that
880 // could have the specified known zero and known one bits, returning them in
882 static void ComputeUnsignedMinMaxValuesFromKnownBits(const Type *Ty,
883 const APInt &KnownZero,
884 const APInt &KnownOne,
885 APInt &Min, APInt &Max) {
886 uint32_t BitWidth = cast<IntegerType>(Ty)->getBitWidth(); BitWidth = BitWidth;
887 assert(KnownZero.getBitWidth() == BitWidth &&
888 KnownOne.getBitWidth() == BitWidth &&
889 Min.getBitWidth() == BitWidth && Max.getBitWidth() &&
890 "Ty, KnownZero, KnownOne and Min, Max must have equal bitwidth.");
891 APInt UnknownBits = ~(KnownZero|KnownOne);
893 // The minimum value is when the unknown bits are all zeros.
895 // The maximum value is when the unknown bits are all ones.
896 Max = KnownOne|UnknownBits;
899 /// SimplifyDemandedBits - This function attempts to replace V with a simpler
900 /// value based on the demanded bits. When this function is called, it is known
901 /// that only the bits set in DemandedMask of the result of V are ever used
902 /// downstream. Consequently, depending on the mask and V, it may be possible
903 /// to replace V with a constant or one of its operands. In such cases, this
904 /// function does the replacement and returns true. In all other cases, it
905 /// returns false after analyzing the expression and setting KnownOne and known
906 /// to be one in the expression. KnownZero contains all the bits that are known
907 /// to be zero in the expression. These are provided to potentially allow the
908 /// caller (which might recursively be SimplifyDemandedBits itself) to simplify
909 /// the expression. KnownOne and KnownZero always follow the invariant that
910 /// KnownOne & KnownZero == 0. That is, a bit can't be both 1 and 0. Note that
911 /// the bits in KnownOne and KnownZero may only be accurate for those bits set
912 /// in DemandedMask. Note also that the bitwidth of V, DemandedMask, KnownZero
913 /// and KnownOne must all be the same.
914 bool InstCombiner::SimplifyDemandedBits(Value *V, APInt DemandedMask,
915 APInt& KnownZero, APInt& KnownOne,
917 assert(V != 0 && "Null pointer of Value???");
918 assert(Depth <= 6 && "Limit Search Depth");
919 uint32_t BitWidth = DemandedMask.getBitWidth();
920 const IntegerType *VTy = cast<IntegerType>(V->getType());
921 assert(VTy->getBitWidth() == BitWidth &&
922 KnownZero.getBitWidth() == BitWidth &&
923 KnownOne.getBitWidth() == BitWidth &&
924 "Value *V, DemandedMask, KnownZero and KnownOne \
925 must have same BitWidth");
926 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
927 // We know all of the bits for a constant!
928 KnownOne = CI->getValue() & DemandedMask;
929 KnownZero = ~KnownOne & DemandedMask;
935 if (!V->hasOneUse()) { // Other users may use these bits.
936 if (Depth != 0) { // Not at the root.
937 // Just compute the KnownZero/KnownOne bits to simplify things downstream.
938 ComputeMaskedBits(V, DemandedMask, KnownZero, KnownOne, Depth);
941 // If this is the root being simplified, allow it to have multiple uses,
942 // just set the DemandedMask to all bits.
943 DemandedMask = APInt::getAllOnesValue(BitWidth);
944 } else if (DemandedMask == 0) { // Not demanding any bits from V.
945 if (V != UndefValue::get(VTy))
946 return UpdateValueUsesWith(V, UndefValue::get(VTy));
948 } else if (Depth == 6) { // Limit search depth.
952 Instruction *I = dyn_cast<Instruction>(V);
953 if (!I) return false; // Only analyze instructions.
955 APInt LHSKnownZero(BitWidth, 0), LHSKnownOne(BitWidth, 0);
956 APInt &RHSKnownZero = KnownZero, &RHSKnownOne = KnownOne;
957 switch (I->getOpcode()) {
959 case Instruction::And:
960 // If either the LHS or the RHS are Zero, the result is zero.
961 if (SimplifyDemandedBits(I->getOperand(1), DemandedMask,
962 RHSKnownZero, RHSKnownOne, Depth+1))
964 assert((RHSKnownZero & RHSKnownOne) == 0 &&
965 "Bits known to be one AND zero?");
967 // If something is known zero on the RHS, the bits aren't demanded on the
969 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask & ~RHSKnownZero,
970 LHSKnownZero, LHSKnownOne, Depth+1))
972 assert((LHSKnownZero & LHSKnownOne) == 0 &&
973 "Bits known to be one AND zero?");
975 // If all of the demanded bits are known 1 on one side, return the other.
976 // These bits cannot contribute to the result of the 'and'.
977 if ((DemandedMask & ~LHSKnownZero & RHSKnownOne) ==
978 (DemandedMask & ~LHSKnownZero))
979 return UpdateValueUsesWith(I, I->getOperand(0));
980 if ((DemandedMask & ~RHSKnownZero & LHSKnownOne) ==
981 (DemandedMask & ~RHSKnownZero))
982 return UpdateValueUsesWith(I, I->getOperand(1));
984 // If all of the demanded bits in the inputs are known zeros, return zero.
985 if ((DemandedMask & (RHSKnownZero|LHSKnownZero)) == DemandedMask)
986 return UpdateValueUsesWith(I, Constant::getNullValue(VTy));
988 // If the RHS is a constant, see if we can simplify it.
989 if (ShrinkDemandedConstant(I, 1, DemandedMask & ~LHSKnownZero))
990 return UpdateValueUsesWith(I, I);
992 // Output known-1 bits are only known if set in both the LHS & RHS.
993 RHSKnownOne &= LHSKnownOne;
994 // Output known-0 are known to be clear if zero in either the LHS | RHS.
995 RHSKnownZero |= LHSKnownZero;
997 case Instruction::Or:
998 // If either the LHS or the RHS are One, the result is One.
999 if (SimplifyDemandedBits(I->getOperand(1), DemandedMask,
1000 RHSKnownZero, RHSKnownOne, Depth+1))
1002 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1003 "Bits known to be one AND zero?");
1004 // If something is known one on the RHS, the bits aren't demanded on the
1006 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask & ~RHSKnownOne,
1007 LHSKnownZero, LHSKnownOne, Depth+1))
1009 assert((LHSKnownZero & LHSKnownOne) == 0 &&
1010 "Bits known to be one AND zero?");
1012 // If all of the demanded bits are known zero on one side, return the other.
1013 // These bits cannot contribute to the result of the 'or'.
1014 if ((DemandedMask & ~LHSKnownOne & RHSKnownZero) ==
1015 (DemandedMask & ~LHSKnownOne))
1016 return UpdateValueUsesWith(I, I->getOperand(0));
1017 if ((DemandedMask & ~RHSKnownOne & LHSKnownZero) ==
1018 (DemandedMask & ~RHSKnownOne))
1019 return UpdateValueUsesWith(I, I->getOperand(1));
1021 // If all of the potentially set bits on one side are known to be set on
1022 // the other side, just use the 'other' side.
1023 if ((DemandedMask & (~RHSKnownZero) & LHSKnownOne) ==
1024 (DemandedMask & (~RHSKnownZero)))
1025 return UpdateValueUsesWith(I, I->getOperand(0));
1026 if ((DemandedMask & (~LHSKnownZero) & RHSKnownOne) ==
1027 (DemandedMask & (~LHSKnownZero)))
1028 return UpdateValueUsesWith(I, I->getOperand(1));
1030 // If the RHS is a constant, see if we can simplify it.
1031 if (ShrinkDemandedConstant(I, 1, DemandedMask))
1032 return UpdateValueUsesWith(I, I);
1034 // Output known-0 bits are only known if clear in both the LHS & RHS.
1035 RHSKnownZero &= LHSKnownZero;
1036 // Output known-1 are known to be set if set in either the LHS | RHS.
1037 RHSKnownOne |= LHSKnownOne;
1039 case Instruction::Xor: {
1040 if (SimplifyDemandedBits(I->getOperand(1), DemandedMask,
1041 RHSKnownZero, RHSKnownOne, Depth+1))
1043 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1044 "Bits known to be one AND zero?");
1045 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask,
1046 LHSKnownZero, LHSKnownOne, Depth+1))
1048 assert((LHSKnownZero & LHSKnownOne) == 0 &&
1049 "Bits known to be one AND zero?");
1051 // If all of the demanded bits are known zero on one side, return the other.
1052 // These bits cannot contribute to the result of the 'xor'.
1053 if ((DemandedMask & RHSKnownZero) == DemandedMask)
1054 return UpdateValueUsesWith(I, I->getOperand(0));
1055 if ((DemandedMask & LHSKnownZero) == DemandedMask)
1056 return UpdateValueUsesWith(I, I->getOperand(1));
1058 // Output known-0 bits are known if clear or set in both the LHS & RHS.
1059 APInt KnownZeroOut = (RHSKnownZero & LHSKnownZero) |
1060 (RHSKnownOne & LHSKnownOne);
1061 // Output known-1 are known to be set if set in only one of the LHS, RHS.
1062 APInt KnownOneOut = (RHSKnownZero & LHSKnownOne) |
1063 (RHSKnownOne & LHSKnownZero);
1065 // If all of the demanded bits are known to be zero on one side or the
1066 // other, turn this into an *inclusive* or.
1067 // e.g. (A & C1)^(B & C2) -> (A & C1)|(B & C2) iff C1&C2 == 0
1068 if ((DemandedMask & ~RHSKnownZero & ~LHSKnownZero) == 0) {
1070 BinaryOperator::createOr(I->getOperand(0), I->getOperand(1),
1072 InsertNewInstBefore(Or, *I);
1073 return UpdateValueUsesWith(I, Or);
1076 // If all of the demanded bits on one side are known, and all of the set
1077 // bits on that side are also known to be set on the other side, turn this
1078 // into an AND, as we know the bits will be cleared.
1079 // e.g. (X | C1) ^ C2 --> (X | C1) & ~C2 iff (C1&C2) == C2
1080 if ((DemandedMask & (RHSKnownZero|RHSKnownOne)) == DemandedMask) {
1082 if ((RHSKnownOne & LHSKnownOne) == RHSKnownOne) {
1083 Constant *AndC = ConstantInt::get(~RHSKnownOne & DemandedMask);
1085 BinaryOperator::createAnd(I->getOperand(0), AndC, "tmp");
1086 InsertNewInstBefore(And, *I);
1087 return UpdateValueUsesWith(I, And);
1091 // If the RHS is a constant, see if we can simplify it.
1092 // FIXME: for XOR, we prefer to force bits to 1 if they will make a -1.
1093 if (ShrinkDemandedConstant(I, 1, DemandedMask))
1094 return UpdateValueUsesWith(I, I);
1096 RHSKnownZero = KnownZeroOut;
1097 RHSKnownOne = KnownOneOut;
1100 case Instruction::Select:
1101 if (SimplifyDemandedBits(I->getOperand(2), DemandedMask,
1102 RHSKnownZero, RHSKnownOne, Depth+1))
1104 if (SimplifyDemandedBits(I->getOperand(1), DemandedMask,
1105 LHSKnownZero, LHSKnownOne, Depth+1))
1107 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1108 "Bits known to be one AND zero?");
1109 assert((LHSKnownZero & LHSKnownOne) == 0 &&
1110 "Bits known to be one AND zero?");
1112 // If the operands are constants, see if we can simplify them.
1113 if (ShrinkDemandedConstant(I, 1, DemandedMask))
1114 return UpdateValueUsesWith(I, I);
1115 if (ShrinkDemandedConstant(I, 2, DemandedMask))
1116 return UpdateValueUsesWith(I, I);
1118 // Only known if known in both the LHS and RHS.
1119 RHSKnownOne &= LHSKnownOne;
1120 RHSKnownZero &= LHSKnownZero;
1122 case Instruction::Trunc: {
1124 cast<IntegerType>(I->getOperand(0)->getType())->getBitWidth();
1125 DemandedMask.zext(truncBf);
1126 RHSKnownZero.zext(truncBf);
1127 RHSKnownOne.zext(truncBf);
1128 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask,
1129 RHSKnownZero, RHSKnownOne, Depth+1))
1131 DemandedMask.trunc(BitWidth);
1132 RHSKnownZero.trunc(BitWidth);
1133 RHSKnownOne.trunc(BitWidth);
1134 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1135 "Bits known to be one AND zero?");
1138 case Instruction::BitCast:
1139 if (!I->getOperand(0)->getType()->isInteger())
1142 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask,
1143 RHSKnownZero, RHSKnownOne, Depth+1))
1145 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1146 "Bits known to be one AND zero?");
1148 case Instruction::ZExt: {
1149 // Compute the bits in the result that are not present in the input.
1150 const IntegerType *SrcTy = cast<IntegerType>(I->getOperand(0)->getType());
1151 uint32_t SrcBitWidth = SrcTy->getBitWidth();
1153 DemandedMask.trunc(SrcBitWidth);
1154 RHSKnownZero.trunc(SrcBitWidth);
1155 RHSKnownOne.trunc(SrcBitWidth);
1156 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask,
1157 RHSKnownZero, RHSKnownOne, Depth+1))
1159 DemandedMask.zext(BitWidth);
1160 RHSKnownZero.zext(BitWidth);
1161 RHSKnownOne.zext(BitWidth);
1162 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1163 "Bits known to be one AND zero?");
1164 // The top bits are known to be zero.
1165 RHSKnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth);
1168 case Instruction::SExt: {
1169 // Compute the bits in the result that are not present in the input.
1170 const IntegerType *SrcTy = cast<IntegerType>(I->getOperand(0)->getType());
1171 uint32_t SrcBitWidth = SrcTy->getBitWidth();
1173 APInt InputDemandedBits = DemandedMask &
1174 APInt::getLowBitsSet(BitWidth, SrcBitWidth);
1176 APInt NewBits(APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth));
1177 // If any of the sign extended bits are demanded, we know that the sign
1179 if ((NewBits & DemandedMask) != 0)
1180 InputDemandedBits.set(SrcBitWidth-1);
1182 InputDemandedBits.trunc(SrcBitWidth);
1183 RHSKnownZero.trunc(SrcBitWidth);
1184 RHSKnownOne.trunc(SrcBitWidth);
1185 if (SimplifyDemandedBits(I->getOperand(0), InputDemandedBits,
1186 RHSKnownZero, RHSKnownOne, Depth+1))
1188 InputDemandedBits.zext(BitWidth);
1189 RHSKnownZero.zext(BitWidth);
1190 RHSKnownOne.zext(BitWidth);
1191 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1192 "Bits known to be one AND zero?");
1194 // If the sign bit of the input is known set or clear, then we know the
1195 // top bits of the result.
1197 // If the input sign bit is known zero, or if the NewBits are not demanded
1198 // convert this into a zero extension.
1199 if (RHSKnownZero[SrcBitWidth-1] || (NewBits & ~DemandedMask) == NewBits)
1201 // Convert to ZExt cast
1202 CastInst *NewCast = new ZExtInst(I->getOperand(0), VTy, I->getName(), I);
1203 return UpdateValueUsesWith(I, NewCast);
1204 } else if (RHSKnownOne[SrcBitWidth-1]) { // Input sign bit known set
1205 RHSKnownOne |= NewBits;
1209 case Instruction::Add: {
1210 // Figure out what the input bits are. If the top bits of the and result
1211 // are not demanded, then the add doesn't demand them from its input
1213 uint32_t NLZ = DemandedMask.countLeadingZeros();
1215 // If there is a constant on the RHS, there are a variety of xformations
1217 if (ConstantInt *RHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
1218 // If null, this should be simplified elsewhere. Some of the xforms here
1219 // won't work if the RHS is zero.
1223 // If the top bit of the output is demanded, demand everything from the
1224 // input. Otherwise, we demand all the input bits except NLZ top bits.
1225 APInt InDemandedBits(APInt::getLowBitsSet(BitWidth, BitWidth - NLZ));
1227 // Find information about known zero/one bits in the input.
1228 if (SimplifyDemandedBits(I->getOperand(0), InDemandedBits,
1229 LHSKnownZero, LHSKnownOne, Depth+1))
1232 // If the RHS of the add has bits set that can't affect the input, reduce
1234 if (ShrinkDemandedConstant(I, 1, InDemandedBits))
1235 return UpdateValueUsesWith(I, I);
1237 // Avoid excess work.
1238 if (LHSKnownZero == 0 && LHSKnownOne == 0)
1241 // Turn it into OR if input bits are zero.
1242 if ((LHSKnownZero & RHS->getValue()) == RHS->getValue()) {
1244 BinaryOperator::createOr(I->getOperand(0), I->getOperand(1),
1246 InsertNewInstBefore(Or, *I);
1247 return UpdateValueUsesWith(I, Or);
1250 // We can say something about the output known-zero and known-one bits,
1251 // depending on potential carries from the input constant and the
1252 // unknowns. For example if the LHS is known to have at most the 0x0F0F0
1253 // bits set and the RHS constant is 0x01001, then we know we have a known
1254 // one mask of 0x00001 and a known zero mask of 0xE0F0E.
1256 // To compute this, we first compute the potential carry bits. These are
1257 // the bits which may be modified. I'm not aware of a better way to do
1259 const APInt& RHSVal = RHS->getValue();
1260 APInt CarryBits((~LHSKnownZero + RHSVal) ^ (~LHSKnownZero ^ RHSVal));
1262 // Now that we know which bits have carries, compute the known-1/0 sets.
1264 // Bits are known one if they are known zero in one operand and one in the
1265 // other, and there is no input carry.
1266 RHSKnownOne = ((LHSKnownZero & RHSVal) |
1267 (LHSKnownOne & ~RHSVal)) & ~CarryBits;
1269 // Bits are known zero if they are known zero in both operands and there
1270 // is no input carry.
1271 RHSKnownZero = LHSKnownZero & ~RHSVal & ~CarryBits;
1273 // If the high-bits of this ADD are not demanded, then it does not demand
1274 // the high bits of its LHS or RHS.
1275 if (DemandedMask[BitWidth-1] == 0) {
1276 // Right fill the mask of bits for this ADD to demand the most
1277 // significant bit and all those below it.
1278 APInt DemandedFromOps(APInt::getLowBitsSet(BitWidth, BitWidth-NLZ));
1279 if (SimplifyDemandedBits(I->getOperand(0), DemandedFromOps,
1280 LHSKnownZero, LHSKnownOne, Depth+1))
1282 if (SimplifyDemandedBits(I->getOperand(1), DemandedFromOps,
1283 LHSKnownZero, LHSKnownOne, Depth+1))
1289 case Instruction::Sub:
1290 // If the high-bits of this SUB are not demanded, then it does not demand
1291 // the high bits of its LHS or RHS.
1292 if (DemandedMask[BitWidth-1] == 0) {
1293 // Right fill the mask of bits for this SUB to demand the most
1294 // significant bit and all those below it.
1295 uint32_t NLZ = DemandedMask.countLeadingZeros();
1296 APInt DemandedFromOps(APInt::getLowBitsSet(BitWidth, BitWidth-NLZ));
1297 if (SimplifyDemandedBits(I->getOperand(0), DemandedFromOps,
1298 LHSKnownZero, LHSKnownOne, Depth+1))
1300 if (SimplifyDemandedBits(I->getOperand(1), DemandedFromOps,
1301 LHSKnownZero, LHSKnownOne, Depth+1))
1305 case Instruction::Shl:
1306 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1307 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth);
1308 APInt DemandedMaskIn(DemandedMask.lshr(ShiftAmt));
1309 if (SimplifyDemandedBits(I->getOperand(0), DemandedMaskIn,
1310 RHSKnownZero, RHSKnownOne, Depth+1))
1312 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1313 "Bits known to be one AND zero?");
1314 RHSKnownZero <<= ShiftAmt;
1315 RHSKnownOne <<= ShiftAmt;
1316 // low bits known zero.
1318 RHSKnownZero |= APInt::getLowBitsSet(BitWidth, ShiftAmt);
1321 case Instruction::LShr:
1322 // For a logical shift right
1323 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1324 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth);
1326 // Unsigned shift right.
1327 APInt DemandedMaskIn(DemandedMask.shl(ShiftAmt));
1328 if (SimplifyDemandedBits(I->getOperand(0), DemandedMaskIn,
1329 RHSKnownZero, RHSKnownOne, Depth+1))
1331 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1332 "Bits known to be one AND zero?");
1333 RHSKnownZero = APIntOps::lshr(RHSKnownZero, ShiftAmt);
1334 RHSKnownOne = APIntOps::lshr(RHSKnownOne, ShiftAmt);
1336 // Compute the new bits that are at the top now.
1337 APInt HighBits(APInt::getHighBitsSet(BitWidth, ShiftAmt));
1338 RHSKnownZero |= HighBits; // high bits known zero.
1342 case Instruction::AShr:
1343 // If this is an arithmetic shift right and only the low-bit is set, we can
1344 // always convert this into a logical shr, even if the shift amount is
1345 // variable. The low bit of the shift cannot be an input sign bit unless
1346 // the shift amount is >= the size of the datatype, which is undefined.
1347 if (DemandedMask == 1) {
1348 // Perform the logical shift right.
1349 Value *NewVal = BinaryOperator::createLShr(
1350 I->getOperand(0), I->getOperand(1), I->getName());
1351 InsertNewInstBefore(cast<Instruction>(NewVal), *I);
1352 return UpdateValueUsesWith(I, NewVal);
1355 // If the sign bit is the only bit demanded by this ashr, then there is no
1356 // need to do it, the shift doesn't change the high bit.
1357 if (DemandedMask.isSignBit())
1358 return UpdateValueUsesWith(I, I->getOperand(0));
1360 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1361 uint32_t ShiftAmt = SA->getLimitedValue(BitWidth);
1363 // Signed shift right.
1364 APInt DemandedMaskIn(DemandedMask.shl(ShiftAmt));
1365 // If any of the "high bits" are demanded, we should set the sign bit as
1367 if (DemandedMask.countLeadingZeros() <= ShiftAmt)
1368 DemandedMaskIn.set(BitWidth-1);
1369 if (SimplifyDemandedBits(I->getOperand(0),
1371 RHSKnownZero, RHSKnownOne, Depth+1))
1373 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1374 "Bits known to be one AND zero?");
1375 // Compute the new bits that are at the top now.
1376 APInt HighBits(APInt::getHighBitsSet(BitWidth, ShiftAmt));
1377 RHSKnownZero = APIntOps::lshr(RHSKnownZero, ShiftAmt);
1378 RHSKnownOne = APIntOps::lshr(RHSKnownOne, ShiftAmt);
1380 // Handle the sign bits.
1381 APInt SignBit(APInt::getSignBit(BitWidth));
1382 // Adjust to where it is now in the mask.
1383 SignBit = APIntOps::lshr(SignBit, ShiftAmt);
1385 // If the input sign bit is known to be zero, or if none of the top bits
1386 // are demanded, turn this into an unsigned shift right.
1387 if (RHSKnownZero[BitWidth-ShiftAmt-1] ||
1388 (HighBits & ~DemandedMask) == HighBits) {
1389 // Perform the logical shift right.
1390 Value *NewVal = BinaryOperator::createLShr(
1391 I->getOperand(0), SA, I->getName());
1392 InsertNewInstBefore(cast<Instruction>(NewVal), *I);
1393 return UpdateValueUsesWith(I, NewVal);
1394 } else if ((RHSKnownOne & SignBit) != 0) { // New bits are known one.
1395 RHSKnownOne |= HighBits;
1401 // If the client is only demanding bits that we know, return the known
1403 if ((DemandedMask & (RHSKnownZero|RHSKnownOne)) == DemandedMask)
1404 return UpdateValueUsesWith(I, ConstantInt::get(RHSKnownOne));
1409 /// SimplifyDemandedVectorElts - The specified value producecs a vector with
1410 /// 64 or fewer elements. DemandedElts contains the set of elements that are
1411 /// actually used by the caller. This method analyzes which elements of the
1412 /// operand are undef and returns that information in UndefElts.
1414 /// If the information about demanded elements can be used to simplify the
1415 /// operation, the operation is simplified, then the resultant value is
1416 /// returned. This returns null if no change was made.
1417 Value *InstCombiner::SimplifyDemandedVectorElts(Value *V, uint64_t DemandedElts,
1418 uint64_t &UndefElts,
1420 unsigned VWidth = cast<VectorType>(V->getType())->getNumElements();
1421 assert(VWidth <= 64 && "Vector too wide to analyze!");
1422 uint64_t EltMask = ~0ULL >> (64-VWidth);
1423 assert(DemandedElts != EltMask && (DemandedElts & ~EltMask) == 0 &&
1424 "Invalid DemandedElts!");
1426 if (isa<UndefValue>(V)) {
1427 // If the entire vector is undefined, just return this info.
1428 UndefElts = EltMask;
1430 } else if (DemandedElts == 0) { // If nothing is demanded, provide undef.
1431 UndefElts = EltMask;
1432 return UndefValue::get(V->getType());
1436 if (ConstantVector *CP = dyn_cast<ConstantVector>(V)) {
1437 const Type *EltTy = cast<VectorType>(V->getType())->getElementType();
1438 Constant *Undef = UndefValue::get(EltTy);
1440 std::vector<Constant*> Elts;
1441 for (unsigned i = 0; i != VWidth; ++i)
1442 if (!(DemandedElts & (1ULL << i))) { // If not demanded, set to undef.
1443 Elts.push_back(Undef);
1444 UndefElts |= (1ULL << i);
1445 } else if (isa<UndefValue>(CP->getOperand(i))) { // Already undef.
1446 Elts.push_back(Undef);
1447 UndefElts |= (1ULL << i);
1448 } else { // Otherwise, defined.
1449 Elts.push_back(CP->getOperand(i));
1452 // If we changed the constant, return it.
1453 Constant *NewCP = ConstantVector::get(Elts);
1454 return NewCP != CP ? NewCP : 0;
1455 } else if (isa<ConstantAggregateZero>(V)) {
1456 // Simplify the CAZ to a ConstantVector where the non-demanded elements are
1458 const Type *EltTy = cast<VectorType>(V->getType())->getElementType();
1459 Constant *Zero = Constant::getNullValue(EltTy);
1460 Constant *Undef = UndefValue::get(EltTy);
1461 std::vector<Constant*> Elts;
1462 for (unsigned i = 0; i != VWidth; ++i)
1463 Elts.push_back((DemandedElts & (1ULL << i)) ? Zero : Undef);
1464 UndefElts = DemandedElts ^ EltMask;
1465 return ConstantVector::get(Elts);
1468 if (!V->hasOneUse()) { // Other users may use these bits.
1469 if (Depth != 0) { // Not at the root.
1470 // TODO: Just compute the UndefElts information recursively.
1474 } else if (Depth == 10) { // Limit search depth.
1478 Instruction *I = dyn_cast<Instruction>(V);
1479 if (!I) return false; // Only analyze instructions.
1481 bool MadeChange = false;
1482 uint64_t UndefElts2;
1484 switch (I->getOpcode()) {
1487 case Instruction::InsertElement: {
1488 // If this is a variable index, we don't know which element it overwrites.
1489 // demand exactly the same input as we produce.
1490 ConstantInt *Idx = dyn_cast<ConstantInt>(I->getOperand(2));
1492 // Note that we can't propagate undef elt info, because we don't know
1493 // which elt is getting updated.
1494 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), DemandedElts,
1495 UndefElts2, Depth+1);
1496 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
1500 // If this is inserting an element that isn't demanded, remove this
1502 unsigned IdxNo = Idx->getZExtValue();
1503 if (IdxNo >= VWidth || (DemandedElts & (1ULL << IdxNo)) == 0)
1504 return AddSoonDeadInstToWorklist(*I, 0);
1506 // Otherwise, the element inserted overwrites whatever was there, so the
1507 // input demanded set is simpler than the output set.
1508 TmpV = SimplifyDemandedVectorElts(I->getOperand(0),
1509 DemandedElts & ~(1ULL << IdxNo),
1510 UndefElts, Depth+1);
1511 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
1513 // The inserted element is defined.
1514 UndefElts |= 1ULL << IdxNo;
1517 case Instruction::BitCast: {
1518 // Vector->vector casts only.
1519 const VectorType *VTy = dyn_cast<VectorType>(I->getOperand(0)->getType());
1521 unsigned InVWidth = VTy->getNumElements();
1522 uint64_t InputDemandedElts = 0;
1525 if (VWidth == InVWidth) {
1526 // If we are converting from <4 x i32> -> <4 x f32>, we demand the same
1527 // elements as are demanded of us.
1529 InputDemandedElts = DemandedElts;
1530 } else if (VWidth > InVWidth) {
1534 // If there are more elements in the result than there are in the source,
1535 // then an input element is live if any of the corresponding output
1536 // elements are live.
1537 Ratio = VWidth/InVWidth;
1538 for (unsigned OutIdx = 0; OutIdx != VWidth; ++OutIdx) {
1539 if (DemandedElts & (1ULL << OutIdx))
1540 InputDemandedElts |= 1ULL << (OutIdx/Ratio);
1546 // If there are more elements in the source than there are in the result,
1547 // then an input element is live if the corresponding output element is
1549 Ratio = InVWidth/VWidth;
1550 for (unsigned InIdx = 0; InIdx != InVWidth; ++InIdx)
1551 if (DemandedElts & (1ULL << InIdx/Ratio))
1552 InputDemandedElts |= 1ULL << InIdx;
1555 // div/rem demand all inputs, because they don't want divide by zero.
1556 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), InputDemandedElts,
1557 UndefElts2, Depth+1);
1559 I->setOperand(0, TmpV);
1563 UndefElts = UndefElts2;
1564 if (VWidth > InVWidth) {
1565 assert(0 && "Unimp");
1566 // If there are more elements in the result than there are in the source,
1567 // then an output element is undef if the corresponding input element is
1569 for (unsigned OutIdx = 0; OutIdx != VWidth; ++OutIdx)
1570 if (UndefElts2 & (1ULL << (OutIdx/Ratio)))
1571 UndefElts |= 1ULL << OutIdx;
1572 } else if (VWidth < InVWidth) {
1573 assert(0 && "Unimp");
1574 // If there are more elements in the source than there are in the result,
1575 // then a result element is undef if all of the corresponding input
1576 // elements are undef.
1577 UndefElts = ~0ULL >> (64-VWidth); // Start out all undef.
1578 for (unsigned InIdx = 0; InIdx != InVWidth; ++InIdx)
1579 if ((UndefElts2 & (1ULL << InIdx)) == 0) // Not undef?
1580 UndefElts &= ~(1ULL << (InIdx/Ratio)); // Clear undef bit.
1584 case Instruction::And:
1585 case Instruction::Or:
1586 case Instruction::Xor:
1587 case Instruction::Add:
1588 case Instruction::Sub:
1589 case Instruction::Mul:
1590 // div/rem demand all inputs, because they don't want divide by zero.
1591 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), DemandedElts,
1592 UndefElts, Depth+1);
1593 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
1594 TmpV = SimplifyDemandedVectorElts(I->getOperand(1), DemandedElts,
1595 UndefElts2, Depth+1);
1596 if (TmpV) { I->setOperand(1, TmpV); MadeChange = true; }
1598 // Output elements are undefined if both are undefined. Consider things
1599 // like undef&0. The result is known zero, not undef.
1600 UndefElts &= UndefElts2;
1603 case Instruction::Call: {
1604 IntrinsicInst *II = dyn_cast<IntrinsicInst>(I);
1606 switch (II->getIntrinsicID()) {
1609 // Binary vector operations that work column-wise. A dest element is a
1610 // function of the corresponding input elements from the two inputs.
1611 case Intrinsic::x86_sse_sub_ss:
1612 case Intrinsic::x86_sse_mul_ss:
1613 case Intrinsic::x86_sse_min_ss:
1614 case Intrinsic::x86_sse_max_ss:
1615 case Intrinsic::x86_sse2_sub_sd:
1616 case Intrinsic::x86_sse2_mul_sd:
1617 case Intrinsic::x86_sse2_min_sd:
1618 case Intrinsic::x86_sse2_max_sd:
1619 TmpV = SimplifyDemandedVectorElts(II->getOperand(1), DemandedElts,
1620 UndefElts, Depth+1);
1621 if (TmpV) { II->setOperand(1, TmpV); MadeChange = true; }
1622 TmpV = SimplifyDemandedVectorElts(II->getOperand(2), DemandedElts,
1623 UndefElts2, Depth+1);
1624 if (TmpV) { II->setOperand(2, TmpV); MadeChange = true; }
1626 // If only the low elt is demanded and this is a scalarizable intrinsic,
1627 // scalarize it now.
1628 if (DemandedElts == 1) {
1629 switch (II->getIntrinsicID()) {
1631 case Intrinsic::x86_sse_sub_ss:
1632 case Intrinsic::x86_sse_mul_ss:
1633 case Intrinsic::x86_sse2_sub_sd:
1634 case Intrinsic::x86_sse2_mul_sd:
1635 // TODO: Lower MIN/MAX/ABS/etc
1636 Value *LHS = II->getOperand(1);
1637 Value *RHS = II->getOperand(2);
1638 // Extract the element as scalars.
1639 LHS = InsertNewInstBefore(new ExtractElementInst(LHS, 0U,"tmp"), *II);
1640 RHS = InsertNewInstBefore(new ExtractElementInst(RHS, 0U,"tmp"), *II);
1642 switch (II->getIntrinsicID()) {
1643 default: assert(0 && "Case stmts out of sync!");
1644 case Intrinsic::x86_sse_sub_ss:
1645 case Intrinsic::x86_sse2_sub_sd:
1646 TmpV = InsertNewInstBefore(BinaryOperator::createSub(LHS, RHS,
1647 II->getName()), *II);
1649 case Intrinsic::x86_sse_mul_ss:
1650 case Intrinsic::x86_sse2_mul_sd:
1651 TmpV = InsertNewInstBefore(BinaryOperator::createMul(LHS, RHS,
1652 II->getName()), *II);
1657 new InsertElementInst(UndefValue::get(II->getType()), TmpV, 0U,
1659 InsertNewInstBefore(New, *II);
1660 AddSoonDeadInstToWorklist(*II, 0);
1665 // Output elements are undefined if both are undefined. Consider things
1666 // like undef&0. The result is known zero, not undef.
1667 UndefElts &= UndefElts2;
1673 return MadeChange ? I : 0;
1676 /// @returns true if the specified compare predicate is
1677 /// true when both operands are equal...
1678 /// @brief Determine if the icmp Predicate is true when both operands are equal
1679 static bool isTrueWhenEqual(ICmpInst::Predicate pred) {
1680 return pred == ICmpInst::ICMP_EQ || pred == ICmpInst::ICMP_UGE ||
1681 pred == ICmpInst::ICMP_SGE || pred == ICmpInst::ICMP_ULE ||
1682 pred == ICmpInst::ICMP_SLE;
1685 /// @returns true if the specified compare instruction is
1686 /// true when both operands are equal...
1687 /// @brief Determine if the ICmpInst returns true when both operands are equal
1688 static bool isTrueWhenEqual(ICmpInst &ICI) {
1689 return isTrueWhenEqual(ICI.getPredicate());
1692 /// AssociativeOpt - Perform an optimization on an associative operator. This
1693 /// function is designed to check a chain of associative operators for a
1694 /// potential to apply a certain optimization. Since the optimization may be
1695 /// applicable if the expression was reassociated, this checks the chain, then
1696 /// reassociates the expression as necessary to expose the optimization
1697 /// opportunity. This makes use of a special Functor, which must define
1698 /// 'shouldApply' and 'apply' methods.
1700 template<typename Functor>
1701 Instruction *AssociativeOpt(BinaryOperator &Root, const Functor &F) {
1702 unsigned Opcode = Root.getOpcode();
1703 Value *LHS = Root.getOperand(0);
1705 // Quick check, see if the immediate LHS matches...
1706 if (F.shouldApply(LHS))
1707 return F.apply(Root);
1709 // Otherwise, if the LHS is not of the same opcode as the root, return.
1710 Instruction *LHSI = dyn_cast<Instruction>(LHS);
1711 while (LHSI && LHSI->getOpcode() == Opcode && LHSI->hasOneUse()) {
1712 // Should we apply this transform to the RHS?
1713 bool ShouldApply = F.shouldApply(LHSI->getOperand(1));
1715 // If not to the RHS, check to see if we should apply to the LHS...
1716 if (!ShouldApply && F.shouldApply(LHSI->getOperand(0))) {
1717 cast<BinaryOperator>(LHSI)->swapOperands(); // Make the LHS the RHS
1721 // If the functor wants to apply the optimization to the RHS of LHSI,
1722 // reassociate the expression from ((? op A) op B) to (? op (A op B))
1724 BasicBlock *BB = Root.getParent();
1726 // Now all of the instructions are in the current basic block, go ahead
1727 // and perform the reassociation.
1728 Instruction *TmpLHSI = cast<Instruction>(Root.getOperand(0));
1730 // First move the selected RHS to the LHS of the root...
1731 Root.setOperand(0, LHSI->getOperand(1));
1733 // Make what used to be the LHS of the root be the user of the root...
1734 Value *ExtraOperand = TmpLHSI->getOperand(1);
1735 if (&Root == TmpLHSI) {
1736 Root.replaceAllUsesWith(Constant::getNullValue(TmpLHSI->getType()));
1739 Root.replaceAllUsesWith(TmpLHSI); // Users now use TmpLHSI
1740 TmpLHSI->setOperand(1, &Root); // TmpLHSI now uses the root
1741 TmpLHSI->getParent()->getInstList().remove(TmpLHSI);
1742 BasicBlock::iterator ARI = &Root; ++ARI;
1743 BB->getInstList().insert(ARI, TmpLHSI); // Move TmpLHSI to after Root
1746 // Now propagate the ExtraOperand down the chain of instructions until we
1748 while (TmpLHSI != LHSI) {
1749 Instruction *NextLHSI = cast<Instruction>(TmpLHSI->getOperand(0));
1750 // Move the instruction to immediately before the chain we are
1751 // constructing to avoid breaking dominance properties.
1752 NextLHSI->getParent()->getInstList().remove(NextLHSI);
1753 BB->getInstList().insert(ARI, NextLHSI);
1756 Value *NextOp = NextLHSI->getOperand(1);
1757 NextLHSI->setOperand(1, ExtraOperand);
1759 ExtraOperand = NextOp;
1762 // Now that the instructions are reassociated, have the functor perform
1763 // the transformation...
1764 return F.apply(Root);
1767 LHSI = dyn_cast<Instruction>(LHSI->getOperand(0));
1773 // AddRHS - Implements: X + X --> X << 1
1776 AddRHS(Value *rhs) : RHS(rhs) {}
1777 bool shouldApply(Value *LHS) const { return LHS == RHS; }
1778 Instruction *apply(BinaryOperator &Add) const {
1779 return BinaryOperator::createShl(Add.getOperand(0),
1780 ConstantInt::get(Add.getType(), 1));
1784 // AddMaskingAnd - Implements (A & C1)+(B & C2) --> (A & C1)|(B & C2)
1786 struct AddMaskingAnd {
1788 AddMaskingAnd(Constant *c) : C2(c) {}
1789 bool shouldApply(Value *LHS) const {
1791 return match(LHS, m_And(m_Value(), m_ConstantInt(C1))) &&
1792 ConstantExpr::getAnd(C1, C2)->isNullValue();
1794 Instruction *apply(BinaryOperator &Add) const {
1795 return BinaryOperator::createOr(Add.getOperand(0), Add.getOperand(1));
1799 static Value *FoldOperationIntoSelectOperand(Instruction &I, Value *SO,
1801 if (CastInst *CI = dyn_cast<CastInst>(&I)) {
1802 if (Constant *SOC = dyn_cast<Constant>(SO))
1803 return ConstantExpr::getCast(CI->getOpcode(), SOC, I.getType());
1805 return IC->InsertNewInstBefore(CastInst::create(
1806 CI->getOpcode(), SO, I.getType(), SO->getName() + ".cast"), I);
1809 // Figure out if the constant is the left or the right argument.
1810 bool ConstIsRHS = isa<Constant>(I.getOperand(1));
1811 Constant *ConstOperand = cast<Constant>(I.getOperand(ConstIsRHS));
1813 if (Constant *SOC = dyn_cast<Constant>(SO)) {
1815 return ConstantExpr::get(I.getOpcode(), SOC, ConstOperand);
1816 return ConstantExpr::get(I.getOpcode(), ConstOperand, SOC);
1819 Value *Op0 = SO, *Op1 = ConstOperand;
1821 std::swap(Op0, Op1);
1823 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(&I))
1824 New = BinaryOperator::create(BO->getOpcode(), Op0, Op1,SO->getName()+".op");
1825 else if (CmpInst *CI = dyn_cast<CmpInst>(&I))
1826 New = CmpInst::create(CI->getOpcode(), CI->getPredicate(), Op0, Op1,
1827 SO->getName()+".cmp");
1829 assert(0 && "Unknown binary instruction type!");
1832 return IC->InsertNewInstBefore(New, I);
1835 // FoldOpIntoSelect - Given an instruction with a select as one operand and a
1836 // constant as the other operand, try to fold the binary operator into the
1837 // select arguments. This also works for Cast instructions, which obviously do
1838 // not have a second operand.
1839 static Instruction *FoldOpIntoSelect(Instruction &Op, SelectInst *SI,
1841 // Don't modify shared select instructions
1842 if (!SI->hasOneUse()) return 0;
1843 Value *TV = SI->getOperand(1);
1844 Value *FV = SI->getOperand(2);
1846 if (isa<Constant>(TV) || isa<Constant>(FV)) {
1847 // Bool selects with constant operands can be folded to logical ops.
1848 if (SI->getType() == Type::Int1Ty) return 0;
1850 Value *SelectTrueVal = FoldOperationIntoSelectOperand(Op, TV, IC);
1851 Value *SelectFalseVal = FoldOperationIntoSelectOperand(Op, FV, IC);
1853 return new SelectInst(SI->getCondition(), SelectTrueVal,
1860 /// FoldOpIntoPhi - Given a binary operator or cast instruction which has a PHI
1861 /// node as operand #0, see if we can fold the instruction into the PHI (which
1862 /// is only possible if all operands to the PHI are constants).
1863 Instruction *InstCombiner::FoldOpIntoPhi(Instruction &I) {
1864 PHINode *PN = cast<PHINode>(I.getOperand(0));
1865 unsigned NumPHIValues = PN->getNumIncomingValues();
1866 if (!PN->hasOneUse() || NumPHIValues == 0) return 0;
1868 // Check to see if all of the operands of the PHI are constants. If there is
1869 // one non-constant value, remember the BB it is. If there is more than one
1870 // or if *it* is a PHI, bail out.
1871 BasicBlock *NonConstBB = 0;
1872 for (unsigned i = 0; i != NumPHIValues; ++i)
1873 if (!isa<Constant>(PN->getIncomingValue(i))) {
1874 if (NonConstBB) return 0; // More than one non-const value.
1875 if (isa<PHINode>(PN->getIncomingValue(i))) return 0; // Itself a phi.
1876 NonConstBB = PN->getIncomingBlock(i);
1878 // If the incoming non-constant value is in I's block, we have an infinite
1880 if (NonConstBB == I.getParent())
1884 // If there is exactly one non-constant value, we can insert a copy of the
1885 // operation in that block. However, if this is a critical edge, we would be
1886 // inserting the computation one some other paths (e.g. inside a loop). Only
1887 // do this if the pred block is unconditionally branching into the phi block.
1889 BranchInst *BI = dyn_cast<BranchInst>(NonConstBB->getTerminator());
1890 if (!BI || !BI->isUnconditional()) return 0;
1893 // Okay, we can do the transformation: create the new PHI node.
1894 PHINode *NewPN = new PHINode(I.getType(), "");
1895 NewPN->reserveOperandSpace(PN->getNumOperands()/2);
1896 InsertNewInstBefore(NewPN, *PN);
1897 NewPN->takeName(PN);
1899 // Next, add all of the operands to the PHI.
1900 if (I.getNumOperands() == 2) {
1901 Constant *C = cast<Constant>(I.getOperand(1));
1902 for (unsigned i = 0; i != NumPHIValues; ++i) {
1904 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i))) {
1905 if (CmpInst *CI = dyn_cast<CmpInst>(&I))
1906 InV = ConstantExpr::getCompare(CI->getPredicate(), InC, C);
1908 InV = ConstantExpr::get(I.getOpcode(), InC, C);
1910 assert(PN->getIncomingBlock(i) == NonConstBB);
1911 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(&I))
1912 InV = BinaryOperator::create(BO->getOpcode(),
1913 PN->getIncomingValue(i), C, "phitmp",
1914 NonConstBB->getTerminator());
1915 else if (CmpInst *CI = dyn_cast<CmpInst>(&I))
1916 InV = CmpInst::create(CI->getOpcode(),
1918 PN->getIncomingValue(i), C, "phitmp",
1919 NonConstBB->getTerminator());
1921 assert(0 && "Unknown binop!");
1923 AddToWorkList(cast<Instruction>(InV));
1925 NewPN->addIncoming(InV, PN->getIncomingBlock(i));
1928 CastInst *CI = cast<CastInst>(&I);
1929 const Type *RetTy = CI->getType();
1930 for (unsigned i = 0; i != NumPHIValues; ++i) {
1932 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i))) {
1933 InV = ConstantExpr::getCast(CI->getOpcode(), InC, RetTy);
1935 assert(PN->getIncomingBlock(i) == NonConstBB);
1936 InV = CastInst::create(CI->getOpcode(), PN->getIncomingValue(i),
1937 I.getType(), "phitmp",
1938 NonConstBB->getTerminator());
1939 AddToWorkList(cast<Instruction>(InV));
1941 NewPN->addIncoming(InV, PN->getIncomingBlock(i));
1944 return ReplaceInstUsesWith(I, NewPN);
1948 /// CannotBeNegativeZero - Return true if we can prove that the specified FP
1949 /// value is never equal to -0.0.
1951 /// Note that this function will need to be revisited when we support nondefault
1954 static bool CannotBeNegativeZero(const Value *V) {
1955 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(V))
1956 return !CFP->getValueAPF().isNegZero();
1958 // (add x, 0.0) is guaranteed to return +0.0, not -0.0.
1959 if (const Instruction *I = dyn_cast<Instruction>(V)) {
1960 if (I->getOpcode() == Instruction::Add &&
1961 isa<ConstantFP>(I->getOperand(1)) &&
1962 cast<ConstantFP>(I->getOperand(1))->isNullValue())
1965 if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(I))
1966 if (II->getIntrinsicID() == Intrinsic::sqrt)
1967 return CannotBeNegativeZero(II->getOperand(1));
1969 if (const CallInst *CI = dyn_cast<CallInst>(I))
1970 if (const Function *F = CI->getCalledFunction()) {
1971 if (F->isDeclaration()) {
1972 switch (F->getNameLen()) {
1973 case 3: // abs(x) != -0.0
1974 if (!strcmp(F->getNameStart(), "abs")) return true;
1976 case 4: // abs[lf](x) != -0.0
1977 if (!strcmp(F->getNameStart(), "absf")) return true;
1978 if (!strcmp(F->getNameStart(), "absl")) return true;
1989 Instruction *InstCombiner::visitAdd(BinaryOperator &I) {
1990 bool Changed = SimplifyCommutative(I);
1991 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
1993 if (Constant *RHSC = dyn_cast<Constant>(RHS)) {
1994 // X + undef -> undef
1995 if (isa<UndefValue>(RHS))
1996 return ReplaceInstUsesWith(I, RHS);
1999 if (!I.getType()->isFPOrFPVector()) { // NOTE: -0 + +0 = +0.
2000 if (RHSC->isNullValue())
2001 return ReplaceInstUsesWith(I, LHS);
2002 } else if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHSC)) {
2003 if (CFP->isExactlyValue(ConstantFP::getNegativeZero
2004 (I.getType())->getValueAPF()))
2005 return ReplaceInstUsesWith(I, LHS);
2008 if (ConstantInt *CI = dyn_cast<ConstantInt>(RHSC)) {
2009 // X + (signbit) --> X ^ signbit
2010 const APInt& Val = CI->getValue();
2011 uint32_t BitWidth = Val.getBitWidth();
2012 if (Val == APInt::getSignBit(BitWidth))
2013 return BinaryOperator::createXor(LHS, RHS);
2015 // See if SimplifyDemandedBits can simplify this. This handles stuff like
2016 // (X & 254)+1 -> (X&254)|1
2017 if (!isa<VectorType>(I.getType())) {
2018 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
2019 if (SimplifyDemandedBits(&I, APInt::getAllOnesValue(BitWidth),
2020 KnownZero, KnownOne))
2025 if (isa<PHINode>(LHS))
2026 if (Instruction *NV = FoldOpIntoPhi(I))
2029 ConstantInt *XorRHS = 0;
2031 if (isa<ConstantInt>(RHSC) &&
2032 match(LHS, m_Xor(m_Value(XorLHS), m_ConstantInt(XorRHS)))) {
2033 uint32_t TySizeBits = I.getType()->getPrimitiveSizeInBits();
2034 const APInt& RHSVal = cast<ConstantInt>(RHSC)->getValue();
2036 uint32_t Size = TySizeBits / 2;
2037 APInt C0080Val(APInt(TySizeBits, 1ULL).shl(Size - 1));
2038 APInt CFF80Val(-C0080Val);
2040 if (TySizeBits > Size) {
2041 // If we have ADD(XOR(AND(X, 0xFF), 0x80), 0xF..F80), it's a sext.
2042 // If we have ADD(XOR(AND(X, 0xFF), 0xF..F80), 0x80), it's a sext.
2043 if ((RHSVal == CFF80Val && XorRHS->getValue() == C0080Val) ||
2044 (RHSVal == C0080Val && XorRHS->getValue() == CFF80Val)) {
2045 // This is a sign extend if the top bits are known zero.
2046 if (!MaskedValueIsZero(XorLHS,
2047 APInt::getHighBitsSet(TySizeBits, TySizeBits - Size)))
2048 Size = 0; // Not a sign ext, but can't be any others either.
2053 C0080Val = APIntOps::lshr(C0080Val, Size);
2054 CFF80Val = APIntOps::ashr(CFF80Val, Size);
2055 } while (Size >= 1);
2057 // FIXME: This shouldn't be necessary. When the backends can handle types
2058 // with funny bit widths then this whole cascade of if statements should
2059 // be removed. It is just here to get the size of the "middle" type back
2060 // up to something that the back ends can handle.
2061 const Type *MiddleType = 0;
2064 case 32: MiddleType = Type::Int32Ty; break;
2065 case 16: MiddleType = Type::Int16Ty; break;
2066 case 8: MiddleType = Type::Int8Ty; break;
2069 Instruction *NewTrunc = new TruncInst(XorLHS, MiddleType, "sext");
2070 InsertNewInstBefore(NewTrunc, I);
2071 return new SExtInst(NewTrunc, I.getType(), I.getName());
2077 if (I.getType()->isInteger() && I.getType() != Type::Int1Ty) {
2078 if (Instruction *Result = AssociativeOpt(I, AddRHS(RHS))) return Result;
2080 if (Instruction *RHSI = dyn_cast<Instruction>(RHS)) {
2081 if (RHSI->getOpcode() == Instruction::Sub)
2082 if (LHS == RHSI->getOperand(1)) // A + (B - A) --> B
2083 return ReplaceInstUsesWith(I, RHSI->getOperand(0));
2085 if (Instruction *LHSI = dyn_cast<Instruction>(LHS)) {
2086 if (LHSI->getOpcode() == Instruction::Sub)
2087 if (RHS == LHSI->getOperand(1)) // (B - A) + A --> B
2088 return ReplaceInstUsesWith(I, LHSI->getOperand(0));
2093 // -A + -B --> -(A + B)
2094 if (Value *LHSV = dyn_castNegVal(LHS)) {
2095 if (Value *RHSV = dyn_castNegVal(RHS)) {
2096 Instruction *NewAdd = BinaryOperator::createAdd(LHSV, RHSV, "sum");
2097 InsertNewInstBefore(NewAdd, I);
2098 return BinaryOperator::createNeg(NewAdd);
2101 return BinaryOperator::createSub(RHS, LHSV);
2105 if (!isa<Constant>(RHS))
2106 if (Value *V = dyn_castNegVal(RHS))
2107 return BinaryOperator::createSub(LHS, V);
2111 if (Value *X = dyn_castFoldableMul(LHS, C2)) {
2112 if (X == RHS) // X*C + X --> X * (C+1)
2113 return BinaryOperator::createMul(RHS, AddOne(C2));
2115 // X*C1 + X*C2 --> X * (C1+C2)
2117 if (X == dyn_castFoldableMul(RHS, C1))
2118 return BinaryOperator::createMul(X, Add(C1, C2));
2121 // X + X*C --> X * (C+1)
2122 if (dyn_castFoldableMul(RHS, C2) == LHS)
2123 return BinaryOperator::createMul(LHS, AddOne(C2));
2125 // X + ~X --> -1 since ~X = -X-1
2126 if (dyn_castNotVal(LHS) == RHS || dyn_castNotVal(RHS) == LHS)
2127 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
2130 // (A & C1)+(B & C2) --> (A & C1)|(B & C2) iff C1&C2 == 0
2131 if (match(RHS, m_And(m_Value(), m_ConstantInt(C2))))
2132 if (Instruction *R = AssociativeOpt(I, AddMaskingAnd(C2)))
2135 // W*X + Y*Z --> W * (X+Z) iff W == Y
2136 if (I.getType()->isIntOrIntVector()) {
2137 Value *W, *X, *Y, *Z;
2138 if (match(LHS, m_Mul(m_Value(W), m_Value(X))) &&
2139 match(RHS, m_Mul(m_Value(Y), m_Value(Z)))) {
2143 } else if (Y == X) {
2145 } else if (X == Z) {
2152 Value *NewAdd = InsertNewInstBefore(BinaryOperator::createAdd(X, Z,
2153 LHS->getName()), I);
2154 return BinaryOperator::createMul(W, NewAdd);
2159 if (ConstantInt *CRHS = dyn_cast<ConstantInt>(RHS)) {
2161 if (match(LHS, m_Not(m_Value(X)))) // ~X + C --> (C-1) - X
2162 return BinaryOperator::createSub(SubOne(CRHS), X);
2164 // (X & FF00) + xx00 -> (X+xx00) & FF00
2165 if (LHS->hasOneUse() && match(LHS, m_And(m_Value(X), m_ConstantInt(C2)))) {
2166 Constant *Anded = And(CRHS, C2);
2167 if (Anded == CRHS) {
2168 // See if all bits from the first bit set in the Add RHS up are included
2169 // in the mask. First, get the rightmost bit.
2170 const APInt& AddRHSV = CRHS->getValue();
2172 // Form a mask of all bits from the lowest bit added through the top.
2173 APInt AddRHSHighBits(~((AddRHSV & -AddRHSV)-1));
2175 // See if the and mask includes all of these bits.
2176 APInt AddRHSHighBitsAnd(AddRHSHighBits & C2->getValue());
2178 if (AddRHSHighBits == AddRHSHighBitsAnd) {
2179 // Okay, the xform is safe. Insert the new add pronto.
2180 Value *NewAdd = InsertNewInstBefore(BinaryOperator::createAdd(X, CRHS,
2181 LHS->getName()), I);
2182 return BinaryOperator::createAnd(NewAdd, C2);
2187 // Try to fold constant add into select arguments.
2188 if (SelectInst *SI = dyn_cast<SelectInst>(LHS))
2189 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2193 // add (cast *A to intptrtype) B ->
2194 // cast (GEP (cast *A to sbyte*) B) --> intptrtype
2196 CastInst *CI = dyn_cast<CastInst>(LHS);
2199 CI = dyn_cast<CastInst>(RHS);
2202 if (CI && CI->getType()->isSized() &&
2203 (CI->getType()->getPrimitiveSizeInBits() ==
2204 TD->getIntPtrType()->getPrimitiveSizeInBits())
2205 && isa<PointerType>(CI->getOperand(0)->getType())) {
2207 cast<PointerType>(CI->getOperand(0)->getType())->getAddressSpace();
2208 Value *I2 = InsertBitCastBefore(CI->getOperand(0),
2209 PointerType::get(Type::Int8Ty, AS), I);
2210 I2 = InsertNewInstBefore(new GetElementPtrInst(I2, Other, "ctg2"), I);
2211 return new PtrToIntInst(I2, CI->getType());
2215 // add (select X 0 (sub n A)) A --> select X A n
2217 SelectInst *SI = dyn_cast<SelectInst>(LHS);
2220 SI = dyn_cast<SelectInst>(RHS);
2223 if (SI && SI->hasOneUse()) {
2224 Value *TV = SI->getTrueValue();
2225 Value *FV = SI->getFalseValue();
2228 // Can we fold the add into the argument of the select?
2229 // We check both true and false select arguments for a matching subtract.
2230 if (match(FV, m_Zero()) && match(TV, m_Sub(m_Value(N), m_Value(A))) &&
2231 A == Other) // Fold the add into the true select value.
2232 return new SelectInst(SI->getCondition(), N, A);
2233 if (match(TV, m_Zero()) && match(FV, m_Sub(m_Value(N), m_Value(A))) &&
2234 A == Other) // Fold the add into the false select value.
2235 return new SelectInst(SI->getCondition(), A, N);
2239 // Check for X+0.0. Simplify it to X if we know X is not -0.0.
2240 if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHS))
2241 if (CFP->getValueAPF().isPosZero() && CannotBeNegativeZero(LHS))
2242 return ReplaceInstUsesWith(I, LHS);
2244 return Changed ? &I : 0;
2247 // isSignBit - Return true if the value represented by the constant only has the
2248 // highest order bit set.
2249 static bool isSignBit(ConstantInt *CI) {
2250 uint32_t NumBits = CI->getType()->getPrimitiveSizeInBits();
2251 return CI->getValue() == APInt::getSignBit(NumBits);
2254 Instruction *InstCombiner::visitSub(BinaryOperator &I) {
2255 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2257 if (Op0 == Op1) // sub X, X -> 0
2258 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2260 // If this is a 'B = x-(-A)', change to B = x+A...
2261 if (Value *V = dyn_castNegVal(Op1))
2262 return BinaryOperator::createAdd(Op0, V);
2264 if (isa<UndefValue>(Op0))
2265 return ReplaceInstUsesWith(I, Op0); // undef - X -> undef
2266 if (isa<UndefValue>(Op1))
2267 return ReplaceInstUsesWith(I, Op1); // X - undef -> undef
2269 if (ConstantInt *C = dyn_cast<ConstantInt>(Op0)) {
2270 // Replace (-1 - A) with (~A)...
2271 if (C->isAllOnesValue())
2272 return BinaryOperator::createNot(Op1);
2274 // C - ~X == X + (1+C)
2276 if (match(Op1, m_Not(m_Value(X))))
2277 return BinaryOperator::createAdd(X, AddOne(C));
2279 // -(X >>u 31) -> (X >>s 31)
2280 // -(X >>s 31) -> (X >>u 31)
2282 if (BinaryOperator *SI = dyn_cast<BinaryOperator>(Op1))
2283 if (SI->getOpcode() == Instruction::LShr) {
2284 if (ConstantInt *CU = dyn_cast<ConstantInt>(SI->getOperand(1))) {
2285 // Check to see if we are shifting out everything but the sign bit.
2286 if (CU->getLimitedValue(SI->getType()->getPrimitiveSizeInBits()) ==
2287 SI->getType()->getPrimitiveSizeInBits()-1) {
2288 // Ok, the transformation is safe. Insert AShr.
2289 return BinaryOperator::create(Instruction::AShr,
2290 SI->getOperand(0), CU, SI->getName());
2294 else if (SI->getOpcode() == Instruction::AShr) {
2295 if (ConstantInt *CU = dyn_cast<ConstantInt>(SI->getOperand(1))) {
2296 // Check to see if we are shifting out everything but the sign bit.
2297 if (CU->getLimitedValue(SI->getType()->getPrimitiveSizeInBits()) ==
2298 SI->getType()->getPrimitiveSizeInBits()-1) {
2299 // Ok, the transformation is safe. Insert LShr.
2300 return BinaryOperator::createLShr(
2301 SI->getOperand(0), CU, SI->getName());
2307 // Try to fold constant sub into select arguments.
2308 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
2309 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2312 if (isa<PHINode>(Op0))
2313 if (Instruction *NV = FoldOpIntoPhi(I))
2317 if (BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1)) {
2318 if (Op1I->getOpcode() == Instruction::Add &&
2319 !Op0->getType()->isFPOrFPVector()) {
2320 if (Op1I->getOperand(0) == Op0) // X-(X+Y) == -Y
2321 return BinaryOperator::createNeg(Op1I->getOperand(1), I.getName());
2322 else if (Op1I->getOperand(1) == Op0) // X-(Y+X) == -Y
2323 return BinaryOperator::createNeg(Op1I->getOperand(0), I.getName());
2324 else if (ConstantInt *CI1 = dyn_cast<ConstantInt>(I.getOperand(0))) {
2325 if (ConstantInt *CI2 = dyn_cast<ConstantInt>(Op1I->getOperand(1)))
2326 // C1-(X+C2) --> (C1-C2)-X
2327 return BinaryOperator::createSub(Subtract(CI1, CI2),
2328 Op1I->getOperand(0));
2332 if (Op1I->hasOneUse()) {
2333 // Replace (x - (y - z)) with (x + (z - y)) if the (y - z) subexpression
2334 // is not used by anyone else...
2336 if (Op1I->getOpcode() == Instruction::Sub &&
2337 !Op1I->getType()->isFPOrFPVector()) {
2338 // Swap the two operands of the subexpr...
2339 Value *IIOp0 = Op1I->getOperand(0), *IIOp1 = Op1I->getOperand(1);
2340 Op1I->setOperand(0, IIOp1);
2341 Op1I->setOperand(1, IIOp0);
2343 // Create the new top level add instruction...
2344 return BinaryOperator::createAdd(Op0, Op1);
2347 // Replace (A - (A & B)) with (A & ~B) if this is the only use of (A&B)...
2349 if (Op1I->getOpcode() == Instruction::And &&
2350 (Op1I->getOperand(0) == Op0 || Op1I->getOperand(1) == Op0)) {
2351 Value *OtherOp = Op1I->getOperand(Op1I->getOperand(0) == Op0);
2354 InsertNewInstBefore(BinaryOperator::createNot(OtherOp, "B.not"), I);
2355 return BinaryOperator::createAnd(Op0, NewNot);
2358 // 0 - (X sdiv C) -> (X sdiv -C)
2359 if (Op1I->getOpcode() == Instruction::SDiv)
2360 if (ConstantInt *CSI = dyn_cast<ConstantInt>(Op0))
2362 if (Constant *DivRHS = dyn_cast<Constant>(Op1I->getOperand(1)))
2363 return BinaryOperator::createSDiv(Op1I->getOperand(0),
2364 ConstantExpr::getNeg(DivRHS));
2366 // X - X*C --> X * (1-C)
2367 ConstantInt *C2 = 0;
2368 if (dyn_castFoldableMul(Op1I, C2) == Op0) {
2369 Constant *CP1 = Subtract(ConstantInt::get(I.getType(), 1), C2);
2370 return BinaryOperator::createMul(Op0, CP1);
2373 // X - ((X / Y) * Y) --> X % Y
2374 if (Op1I->getOpcode() == Instruction::Mul)
2375 if (Instruction *I = dyn_cast<Instruction>(Op1I->getOperand(0)))
2376 if (Op0 == I->getOperand(0) &&
2377 Op1I->getOperand(1) == I->getOperand(1)) {
2378 if (I->getOpcode() == Instruction::SDiv)
2379 return BinaryOperator::createSRem(Op0, Op1I->getOperand(1));
2380 if (I->getOpcode() == Instruction::UDiv)
2381 return BinaryOperator::createURem(Op0, Op1I->getOperand(1));
2386 if (!Op0->getType()->isFPOrFPVector())
2387 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0))
2388 if (Op0I->getOpcode() == Instruction::Add) {
2389 if (Op0I->getOperand(0) == Op1) // (Y+X)-Y == X
2390 return ReplaceInstUsesWith(I, Op0I->getOperand(1));
2391 else if (Op0I->getOperand(1) == Op1) // (X+Y)-Y == X
2392 return ReplaceInstUsesWith(I, Op0I->getOperand(0));
2393 } else if (Op0I->getOpcode() == Instruction::Sub) {
2394 if (Op0I->getOperand(0) == Op1) // (X-Y)-X == -Y
2395 return BinaryOperator::createNeg(Op0I->getOperand(1), I.getName());
2399 if (Value *X = dyn_castFoldableMul(Op0, C1)) {
2400 if (X == Op1) // X*C - X --> X * (C-1)
2401 return BinaryOperator::createMul(Op1, SubOne(C1));
2403 ConstantInt *C2; // X*C1 - X*C2 -> X * (C1-C2)
2404 if (X == dyn_castFoldableMul(Op1, C2))
2405 return BinaryOperator::createMul(Op1, Subtract(C1, C2));
2410 /// isSignBitCheck - Given an exploded icmp instruction, return true if the
2411 /// comparison only checks the sign bit. If it only checks the sign bit, set
2412 /// TrueIfSigned if the result of the comparison is true when the input value is
2414 static bool isSignBitCheck(ICmpInst::Predicate pred, ConstantInt *RHS,
2415 bool &TrueIfSigned) {
2417 case ICmpInst::ICMP_SLT: // True if LHS s< 0
2418 TrueIfSigned = true;
2419 return RHS->isZero();
2420 case ICmpInst::ICMP_SLE: // True if LHS s<= RHS and RHS == -1
2421 TrueIfSigned = true;
2422 return RHS->isAllOnesValue();
2423 case ICmpInst::ICMP_SGT: // True if LHS s> -1
2424 TrueIfSigned = false;
2425 return RHS->isAllOnesValue();
2426 case ICmpInst::ICMP_UGT:
2427 // True if LHS u> RHS and RHS == high-bit-mask - 1
2428 TrueIfSigned = true;
2429 return RHS->getValue() ==
2430 APInt::getSignedMaxValue(RHS->getType()->getPrimitiveSizeInBits());
2431 case ICmpInst::ICMP_UGE:
2432 // True if LHS u>= RHS and RHS == high-bit-mask (2^7, 2^15, 2^31, etc)
2433 TrueIfSigned = true;
2434 return RHS->getValue() ==
2435 APInt::getSignBit(RHS->getType()->getPrimitiveSizeInBits());
2441 Instruction *InstCombiner::visitMul(BinaryOperator &I) {
2442 bool Changed = SimplifyCommutative(I);
2443 Value *Op0 = I.getOperand(0);
2445 if (isa<UndefValue>(I.getOperand(1))) // undef * X -> 0
2446 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2448 // Simplify mul instructions with a constant RHS...
2449 if (Constant *Op1 = dyn_cast<Constant>(I.getOperand(1))) {
2450 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
2452 // ((X << C1)*C2) == (X * (C2 << C1))
2453 if (BinaryOperator *SI = dyn_cast<BinaryOperator>(Op0))
2454 if (SI->getOpcode() == Instruction::Shl)
2455 if (Constant *ShOp = dyn_cast<Constant>(SI->getOperand(1)))
2456 return BinaryOperator::createMul(SI->getOperand(0),
2457 ConstantExpr::getShl(CI, ShOp));
2460 return ReplaceInstUsesWith(I, Op1); // X * 0 == 0
2461 if (CI->equalsInt(1)) // X * 1 == X
2462 return ReplaceInstUsesWith(I, Op0);
2463 if (CI->isAllOnesValue()) // X * -1 == 0 - X
2464 return BinaryOperator::createNeg(Op0, I.getName());
2466 const APInt& Val = cast<ConstantInt>(CI)->getValue();
2467 if (Val.isPowerOf2()) { // Replace X*(2^C) with X << C
2468 return BinaryOperator::createShl(Op0,
2469 ConstantInt::get(Op0->getType(), Val.logBase2()));
2471 } else if (ConstantFP *Op1F = dyn_cast<ConstantFP>(Op1)) {
2472 if (Op1F->isNullValue())
2473 return ReplaceInstUsesWith(I, Op1);
2475 // "In IEEE floating point, x*1 is not equivalent to x for nans. However,
2476 // ANSI says we can drop signals, so we can do this anyway." (from GCC)
2477 // We need a better interface for long double here.
2478 if (Op1->getType() == Type::FloatTy || Op1->getType() == Type::DoubleTy)
2479 if (Op1F->isExactlyValue(1.0))
2480 return ReplaceInstUsesWith(I, Op0); // Eliminate 'mul double %X, 1.0'
2483 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0))
2484 if (Op0I->getOpcode() == Instruction::Add && Op0I->hasOneUse() &&
2485 isa<ConstantInt>(Op0I->getOperand(1))) {
2486 // Canonicalize (X+C1)*C2 -> X*C2+C1*C2.
2487 Instruction *Add = BinaryOperator::createMul(Op0I->getOperand(0),
2489 InsertNewInstBefore(Add, I);
2490 Value *C1C2 = ConstantExpr::getMul(Op1,
2491 cast<Constant>(Op0I->getOperand(1)));
2492 return BinaryOperator::createAdd(Add, C1C2);
2496 // Try to fold constant mul into select arguments.
2497 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
2498 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2501 if (isa<PHINode>(Op0))
2502 if (Instruction *NV = FoldOpIntoPhi(I))
2506 if (Value *Op0v = dyn_castNegVal(Op0)) // -X * -Y = X*Y
2507 if (Value *Op1v = dyn_castNegVal(I.getOperand(1)))
2508 return BinaryOperator::createMul(Op0v, Op1v);
2510 // If one of the operands of the multiply is a cast from a boolean value, then
2511 // we know the bool is either zero or one, so this is a 'masking' multiply.
2512 // See if we can simplify things based on how the boolean was originally
2514 CastInst *BoolCast = 0;
2515 if (ZExtInst *CI = dyn_cast<ZExtInst>(I.getOperand(0)))
2516 if (CI->getOperand(0)->getType() == Type::Int1Ty)
2519 if (ZExtInst *CI = dyn_cast<ZExtInst>(I.getOperand(1)))
2520 if (CI->getOperand(0)->getType() == Type::Int1Ty)
2523 if (ICmpInst *SCI = dyn_cast<ICmpInst>(BoolCast->getOperand(0))) {
2524 Value *SCIOp0 = SCI->getOperand(0), *SCIOp1 = SCI->getOperand(1);
2525 const Type *SCOpTy = SCIOp0->getType();
2528 // If the icmp is true iff the sign bit of X is set, then convert this
2529 // multiply into a shift/and combination.
2530 if (isa<ConstantInt>(SCIOp1) &&
2531 isSignBitCheck(SCI->getPredicate(), cast<ConstantInt>(SCIOp1), TIS) &&
2533 // Shift the X value right to turn it into "all signbits".
2534 Constant *Amt = ConstantInt::get(SCIOp0->getType(),
2535 SCOpTy->getPrimitiveSizeInBits()-1);
2537 InsertNewInstBefore(
2538 BinaryOperator::create(Instruction::AShr, SCIOp0, Amt,
2539 BoolCast->getOperand(0)->getName()+
2542 // If the multiply type is not the same as the source type, sign extend
2543 // or truncate to the multiply type.
2544 if (I.getType() != V->getType()) {
2545 uint32_t SrcBits = V->getType()->getPrimitiveSizeInBits();
2546 uint32_t DstBits = I.getType()->getPrimitiveSizeInBits();
2547 Instruction::CastOps opcode =
2548 (SrcBits == DstBits ? Instruction::BitCast :
2549 (SrcBits < DstBits ? Instruction::SExt : Instruction::Trunc));
2550 V = InsertCastBefore(opcode, V, I.getType(), I);
2553 Value *OtherOp = Op0 == BoolCast ? I.getOperand(1) : Op0;
2554 return BinaryOperator::createAnd(V, OtherOp);
2559 return Changed ? &I : 0;
2562 /// This function implements the transforms on div instructions that work
2563 /// regardless of the kind of div instruction it is (udiv, sdiv, or fdiv). It is
2564 /// used by the visitors to those instructions.
2565 /// @brief Transforms common to all three div instructions
2566 Instruction *InstCombiner::commonDivTransforms(BinaryOperator &I) {
2567 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2570 if (isa<UndefValue>(Op0))
2571 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2573 // X / undef -> undef
2574 if (isa<UndefValue>(Op1))
2575 return ReplaceInstUsesWith(I, Op1);
2577 // Handle cases involving: [su]div X, (select Cond, Y, Z)
2578 // This does not apply for fdiv.
2579 if (SelectInst *SI = dyn_cast<SelectInst>(Op1)) {
2580 // [su]div X, (Cond ? 0 : Y) -> div X, Y. If the div and the select are in
2581 // the same basic block, then we replace the select with Y, and the
2582 // condition of the select with false (if the cond value is in the same BB).
2583 // If the select has uses other than the div, this allows them to be
2584 // simplified also. Note that div X, Y is just as good as div X, 0 (undef)
2585 if (ConstantInt *ST = dyn_cast<ConstantInt>(SI->getOperand(1)))
2586 if (ST->isNullValue()) {
2587 Instruction *CondI = dyn_cast<Instruction>(SI->getOperand(0));
2588 if (CondI && CondI->getParent() == I.getParent())
2589 UpdateValueUsesWith(CondI, ConstantInt::getFalse());
2590 else if (I.getParent() != SI->getParent() || SI->hasOneUse())
2591 I.setOperand(1, SI->getOperand(2));
2593 UpdateValueUsesWith(SI, SI->getOperand(2));
2597 // Likewise for: [su]div X, (Cond ? Y : 0) -> div X, Y
2598 if (ConstantInt *ST = dyn_cast<ConstantInt>(SI->getOperand(2)))
2599 if (ST->isNullValue()) {
2600 Instruction *CondI = dyn_cast<Instruction>(SI->getOperand(0));
2601 if (CondI && CondI->getParent() == I.getParent())
2602 UpdateValueUsesWith(CondI, ConstantInt::getTrue());
2603 else if (I.getParent() != SI->getParent() || SI->hasOneUse())
2604 I.setOperand(1, SI->getOperand(1));
2606 UpdateValueUsesWith(SI, SI->getOperand(1));
2614 /// This function implements the transforms common to both integer division
2615 /// instructions (udiv and sdiv). It is called by the visitors to those integer
2616 /// division instructions.
2617 /// @brief Common integer divide transforms
2618 Instruction *InstCombiner::commonIDivTransforms(BinaryOperator &I) {
2619 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2621 if (Instruction *Common = commonDivTransforms(I))
2624 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
2626 if (RHS->equalsInt(1))
2627 return ReplaceInstUsesWith(I, Op0);
2629 // (X / C1) / C2 -> X / (C1*C2)
2630 if (Instruction *LHS = dyn_cast<Instruction>(Op0))
2631 if (Instruction::BinaryOps(LHS->getOpcode()) == I.getOpcode())
2632 if (ConstantInt *LHSRHS = dyn_cast<ConstantInt>(LHS->getOperand(1))) {
2633 return BinaryOperator::create(I.getOpcode(), LHS->getOperand(0),
2634 Multiply(RHS, LHSRHS));
2637 if (!RHS->isZero()) { // avoid X udiv 0
2638 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
2639 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2641 if (isa<PHINode>(Op0))
2642 if (Instruction *NV = FoldOpIntoPhi(I))
2647 // 0 / X == 0, we don't need to preserve faults!
2648 if (ConstantInt *LHS = dyn_cast<ConstantInt>(Op0))
2649 if (LHS->equalsInt(0))
2650 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2655 Instruction *InstCombiner::visitUDiv(BinaryOperator &I) {
2656 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2658 // Handle the integer div common cases
2659 if (Instruction *Common = commonIDivTransforms(I))
2662 // X udiv C^2 -> X >> C
2663 // Check to see if this is an unsigned division with an exact power of 2,
2664 // if so, convert to a right shift.
2665 if (ConstantInt *C = dyn_cast<ConstantInt>(Op1)) {
2666 if (C->getValue().isPowerOf2()) // 0 not included in isPowerOf2
2667 return BinaryOperator::createLShr(Op0,
2668 ConstantInt::get(Op0->getType(), C->getValue().logBase2()));
2671 // X udiv (C1 << N), where C1 is "1<<C2" --> X >> (N+C2)
2672 if (BinaryOperator *RHSI = dyn_cast<BinaryOperator>(I.getOperand(1))) {
2673 if (RHSI->getOpcode() == Instruction::Shl &&
2674 isa<ConstantInt>(RHSI->getOperand(0))) {
2675 const APInt& C1 = cast<ConstantInt>(RHSI->getOperand(0))->getValue();
2676 if (C1.isPowerOf2()) {
2677 Value *N = RHSI->getOperand(1);
2678 const Type *NTy = N->getType();
2679 if (uint32_t C2 = C1.logBase2()) {
2680 Constant *C2V = ConstantInt::get(NTy, C2);
2681 N = InsertNewInstBefore(BinaryOperator::createAdd(N, C2V, "tmp"), I);
2683 return BinaryOperator::createLShr(Op0, N);
2688 // udiv X, (Select Cond, C1, C2) --> Select Cond, (shr X, C1), (shr X, C2)
2689 // where C1&C2 are powers of two.
2690 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
2691 if (ConstantInt *STO = dyn_cast<ConstantInt>(SI->getOperand(1)))
2692 if (ConstantInt *SFO = dyn_cast<ConstantInt>(SI->getOperand(2))) {
2693 const APInt &TVA = STO->getValue(), &FVA = SFO->getValue();
2694 if (TVA.isPowerOf2() && FVA.isPowerOf2()) {
2695 // Compute the shift amounts
2696 uint32_t TSA = TVA.logBase2(), FSA = FVA.logBase2();
2697 // Construct the "on true" case of the select
2698 Constant *TC = ConstantInt::get(Op0->getType(), TSA);
2699 Instruction *TSI = BinaryOperator::createLShr(
2700 Op0, TC, SI->getName()+".t");
2701 TSI = InsertNewInstBefore(TSI, I);
2703 // Construct the "on false" case of the select
2704 Constant *FC = ConstantInt::get(Op0->getType(), FSA);
2705 Instruction *FSI = BinaryOperator::createLShr(
2706 Op0, FC, SI->getName()+".f");
2707 FSI = InsertNewInstBefore(FSI, I);
2709 // construct the select instruction and return it.
2710 return new SelectInst(SI->getOperand(0), TSI, FSI, SI->getName());
2716 Instruction *InstCombiner::visitSDiv(BinaryOperator &I) {
2717 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2719 // Handle the integer div common cases
2720 if (Instruction *Common = commonIDivTransforms(I))
2723 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
2725 if (RHS->isAllOnesValue())
2726 return BinaryOperator::createNeg(Op0);
2729 if (Value *LHSNeg = dyn_castNegVal(Op0))
2730 return BinaryOperator::createSDiv(LHSNeg, ConstantExpr::getNeg(RHS));
2733 // If the sign bits of both operands are zero (i.e. we can prove they are
2734 // unsigned inputs), turn this into a udiv.
2735 if (I.getType()->isInteger()) {
2736 APInt Mask(APInt::getSignBit(I.getType()->getPrimitiveSizeInBits()));
2737 if (MaskedValueIsZero(Op1, Mask) && MaskedValueIsZero(Op0, Mask)) {
2738 // X sdiv Y -> X udiv Y, iff X and Y don't have sign bit set
2739 return BinaryOperator::createUDiv(Op0, Op1, I.getName());
2746 Instruction *InstCombiner::visitFDiv(BinaryOperator &I) {
2747 return commonDivTransforms(I);
2750 /// GetFactor - If we can prove that the specified value is at least a multiple
2751 /// of some factor, return that factor.
2752 static Constant *GetFactor(Value *V) {
2753 if (ConstantInt *CI = dyn_cast<ConstantInt>(V))
2756 // Unless we can be tricky, we know this is a multiple of 1.
2757 Constant *Result = ConstantInt::get(V->getType(), 1);
2759 Instruction *I = dyn_cast<Instruction>(V);
2760 if (!I) return Result;
2762 if (I->getOpcode() == Instruction::Mul) {
2763 // Handle multiplies by a constant, etc.
2764 return ConstantExpr::getMul(GetFactor(I->getOperand(0)),
2765 GetFactor(I->getOperand(1)));
2766 } else if (I->getOpcode() == Instruction::Shl) {
2767 // (X<<C) -> X * (1 << C)
2768 if (Constant *ShRHS = dyn_cast<Constant>(I->getOperand(1))) {
2769 ShRHS = ConstantExpr::getShl(Result, ShRHS);
2770 return ConstantExpr::getMul(GetFactor(I->getOperand(0)), ShRHS);
2772 } else if (I->getOpcode() == Instruction::And) {
2773 if (ConstantInt *RHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
2774 // X & 0xFFF0 is known to be a multiple of 16.
2775 uint32_t Zeros = RHS->getValue().countTrailingZeros();
2776 if (Zeros != V->getType()->getPrimitiveSizeInBits())// don't shift by "32"
2777 return ConstantExpr::getShl(Result,
2778 ConstantInt::get(Result->getType(), Zeros));
2780 } else if (CastInst *CI = dyn_cast<CastInst>(I)) {
2781 // Only handle int->int casts.
2782 if (!CI->isIntegerCast())
2784 Value *Op = CI->getOperand(0);
2785 return ConstantExpr::getCast(CI->getOpcode(), GetFactor(Op), V->getType());
2790 /// This function implements the transforms on rem instructions that work
2791 /// regardless of the kind of rem instruction it is (urem, srem, or frem). It
2792 /// is used by the visitors to those instructions.
2793 /// @brief Transforms common to all three rem instructions
2794 Instruction *InstCombiner::commonRemTransforms(BinaryOperator &I) {
2795 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2797 // 0 % X == 0, we don't need to preserve faults!
2798 if (Constant *LHS = dyn_cast<Constant>(Op0))
2799 if (LHS->isNullValue())
2800 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2802 if (isa<UndefValue>(Op0)) // undef % X -> 0
2803 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2804 if (isa<UndefValue>(Op1))
2805 return ReplaceInstUsesWith(I, Op1); // X % undef -> undef
2807 // Handle cases involving: rem X, (select Cond, Y, Z)
2808 if (SelectInst *SI = dyn_cast<SelectInst>(Op1)) {
2809 // rem X, (Cond ? 0 : Y) -> rem X, Y. If the rem and the select are in
2810 // the same basic block, then we replace the select with Y, and the
2811 // condition of the select with false (if the cond value is in the same
2812 // BB). If the select has uses other than the div, this allows them to be
2814 if (Constant *ST = dyn_cast<Constant>(SI->getOperand(1)))
2815 if (ST->isNullValue()) {
2816 Instruction *CondI = dyn_cast<Instruction>(SI->getOperand(0));
2817 if (CondI && CondI->getParent() == I.getParent())
2818 UpdateValueUsesWith(CondI, ConstantInt::getFalse());
2819 else if (I.getParent() != SI->getParent() || SI->hasOneUse())
2820 I.setOperand(1, SI->getOperand(2));
2822 UpdateValueUsesWith(SI, SI->getOperand(2));
2825 // Likewise for: rem X, (Cond ? Y : 0) -> rem X, Y
2826 if (Constant *ST = dyn_cast<Constant>(SI->getOperand(2)))
2827 if (ST->isNullValue()) {
2828 Instruction *CondI = dyn_cast<Instruction>(SI->getOperand(0));
2829 if (CondI && CondI->getParent() == I.getParent())
2830 UpdateValueUsesWith(CondI, ConstantInt::getTrue());
2831 else if (I.getParent() != SI->getParent() || SI->hasOneUse())
2832 I.setOperand(1, SI->getOperand(1));
2834 UpdateValueUsesWith(SI, SI->getOperand(1));
2842 /// This function implements the transforms common to both integer remainder
2843 /// instructions (urem and srem). It is called by the visitors to those integer
2844 /// remainder instructions.
2845 /// @brief Common integer remainder transforms
2846 Instruction *InstCombiner::commonIRemTransforms(BinaryOperator &I) {
2847 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2849 if (Instruction *common = commonRemTransforms(I))
2852 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
2853 // X % 0 == undef, we don't need to preserve faults!
2854 if (RHS->equalsInt(0))
2855 return ReplaceInstUsesWith(I, UndefValue::get(I.getType()));
2857 if (RHS->equalsInt(1)) // X % 1 == 0
2858 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2860 if (Instruction *Op0I = dyn_cast<Instruction>(Op0)) {
2861 if (SelectInst *SI = dyn_cast<SelectInst>(Op0I)) {
2862 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2864 } else if (isa<PHINode>(Op0I)) {
2865 if (Instruction *NV = FoldOpIntoPhi(I))
2868 // (X * C1) % C2 --> 0 iff C1 % C2 == 0
2869 if (ConstantExpr::getSRem(GetFactor(Op0I), RHS)->isNullValue())
2870 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2877 Instruction *InstCombiner::visitURem(BinaryOperator &I) {
2878 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2880 if (Instruction *common = commonIRemTransforms(I))
2883 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
2884 // X urem C^2 -> X and C
2885 // Check to see if this is an unsigned remainder with an exact power of 2,
2886 // if so, convert to a bitwise and.
2887 if (ConstantInt *C = dyn_cast<ConstantInt>(RHS))
2888 if (C->getValue().isPowerOf2())
2889 return BinaryOperator::createAnd(Op0, SubOne(C));
2892 if (Instruction *RHSI = dyn_cast<Instruction>(I.getOperand(1))) {
2893 // Turn A % (C << N), where C is 2^k, into A & ((C << N)-1)
2894 if (RHSI->getOpcode() == Instruction::Shl &&
2895 isa<ConstantInt>(RHSI->getOperand(0))) {
2896 if (cast<ConstantInt>(RHSI->getOperand(0))->getValue().isPowerOf2()) {
2897 Constant *N1 = ConstantInt::getAllOnesValue(I.getType());
2898 Value *Add = InsertNewInstBefore(BinaryOperator::createAdd(RHSI, N1,
2900 return BinaryOperator::createAnd(Op0, Add);
2905 // urem X, (select Cond, 2^C1, 2^C2) --> select Cond, (and X, C1), (and X, C2)
2906 // where C1&C2 are powers of two.
2907 if (SelectInst *SI = dyn_cast<SelectInst>(Op1)) {
2908 if (ConstantInt *STO = dyn_cast<ConstantInt>(SI->getOperand(1)))
2909 if (ConstantInt *SFO = dyn_cast<ConstantInt>(SI->getOperand(2))) {
2910 // STO == 0 and SFO == 0 handled above.
2911 if ((STO->getValue().isPowerOf2()) &&
2912 (SFO->getValue().isPowerOf2())) {
2913 Value *TrueAnd = InsertNewInstBefore(
2914 BinaryOperator::createAnd(Op0, SubOne(STO), SI->getName()+".t"), I);
2915 Value *FalseAnd = InsertNewInstBefore(
2916 BinaryOperator::createAnd(Op0, SubOne(SFO), SI->getName()+".f"), I);
2917 return new SelectInst(SI->getOperand(0), TrueAnd, FalseAnd);
2925 Instruction *InstCombiner::visitSRem(BinaryOperator &I) {
2926 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2928 // Handle the integer rem common cases
2929 if (Instruction *common = commonIRemTransforms(I))
2932 if (Value *RHSNeg = dyn_castNegVal(Op1))
2933 if (!isa<ConstantInt>(RHSNeg) ||
2934 cast<ConstantInt>(RHSNeg)->getValue().isStrictlyPositive()) {
2936 AddUsesToWorkList(I);
2937 I.setOperand(1, RHSNeg);
2941 // If the sign bits of both operands are zero (i.e. we can prove they are
2942 // unsigned inputs), turn this into a urem.
2943 if (I.getType()->isInteger()) {
2944 APInt Mask(APInt::getSignBit(I.getType()->getPrimitiveSizeInBits()));
2945 if (MaskedValueIsZero(Op1, Mask) && MaskedValueIsZero(Op0, Mask)) {
2946 // X srem Y -> X urem Y, iff X and Y don't have sign bit set
2947 return BinaryOperator::createURem(Op0, Op1, I.getName());
2954 Instruction *InstCombiner::visitFRem(BinaryOperator &I) {
2955 return commonRemTransforms(I);
2958 // isMaxValueMinusOne - return true if this is Max-1
2959 static bool isMaxValueMinusOne(const ConstantInt *C, bool isSigned) {
2960 uint32_t TypeBits = C->getType()->getPrimitiveSizeInBits();
2962 return C->getValue() == APInt::getAllOnesValue(TypeBits) - 1;
2963 return C->getValue() == APInt::getSignedMaxValue(TypeBits)-1;
2966 // isMinValuePlusOne - return true if this is Min+1
2967 static bool isMinValuePlusOne(const ConstantInt *C, bool isSigned) {
2969 return C->getValue() == 1; // unsigned
2971 // Calculate 1111111111000000000000
2972 uint32_t TypeBits = C->getType()->getPrimitiveSizeInBits();
2973 return C->getValue() == APInt::getSignedMinValue(TypeBits)+1;
2976 // isOneBitSet - Return true if there is exactly one bit set in the specified
2978 static bool isOneBitSet(const ConstantInt *CI) {
2979 return CI->getValue().isPowerOf2();
2982 // isHighOnes - Return true if the constant is of the form 1+0+.
2983 // This is the same as lowones(~X).
2984 static bool isHighOnes(const ConstantInt *CI) {
2985 return (~CI->getValue() + 1).isPowerOf2();
2988 /// getICmpCode - Encode a icmp predicate into a three bit mask. These bits
2989 /// are carefully arranged to allow folding of expressions such as:
2991 /// (A < B) | (A > B) --> (A != B)
2993 /// Note that this is only valid if the first and second predicates have the
2994 /// same sign. Is illegal to do: (A u< B) | (A s> B)
2996 /// Three bits are used to represent the condition, as follows:
3001 /// <=> Value Definition
3002 /// 000 0 Always false
3009 /// 111 7 Always true
3011 static unsigned getICmpCode(const ICmpInst *ICI) {
3012 switch (ICI->getPredicate()) {
3014 case ICmpInst::ICMP_UGT: return 1; // 001
3015 case ICmpInst::ICMP_SGT: return 1; // 001
3016 case ICmpInst::ICMP_EQ: return 2; // 010
3017 case ICmpInst::ICMP_UGE: return 3; // 011
3018 case ICmpInst::ICMP_SGE: return 3; // 011
3019 case ICmpInst::ICMP_ULT: return 4; // 100
3020 case ICmpInst::ICMP_SLT: return 4; // 100
3021 case ICmpInst::ICMP_NE: return 5; // 101
3022 case ICmpInst::ICMP_ULE: return 6; // 110
3023 case ICmpInst::ICMP_SLE: return 6; // 110
3026 assert(0 && "Invalid ICmp predicate!");
3031 /// getICmpValue - This is the complement of getICmpCode, which turns an
3032 /// opcode and two operands into either a constant true or false, or a brand
3033 /// new ICmp instruction. The sign is passed in to determine which kind
3034 /// of predicate to use in new icmp instructions.
3035 static Value *getICmpValue(bool sign, unsigned code, Value *LHS, Value *RHS) {
3037 default: assert(0 && "Illegal ICmp code!");
3038 case 0: return ConstantInt::getFalse();
3041 return new ICmpInst(ICmpInst::ICMP_SGT, LHS, RHS);
3043 return new ICmpInst(ICmpInst::ICMP_UGT, LHS, RHS);
3044 case 2: return new ICmpInst(ICmpInst::ICMP_EQ, LHS, RHS);
3047 return new ICmpInst(ICmpInst::ICMP_SGE, LHS, RHS);
3049 return new ICmpInst(ICmpInst::ICMP_UGE, LHS, RHS);
3052 return new ICmpInst(ICmpInst::ICMP_SLT, LHS, RHS);
3054 return new ICmpInst(ICmpInst::ICMP_ULT, LHS, RHS);
3055 case 5: return new ICmpInst(ICmpInst::ICMP_NE, LHS, RHS);
3058 return new ICmpInst(ICmpInst::ICMP_SLE, LHS, RHS);
3060 return new ICmpInst(ICmpInst::ICMP_ULE, LHS, RHS);
3061 case 7: return ConstantInt::getTrue();
3065 static bool PredicatesFoldable(ICmpInst::Predicate p1, ICmpInst::Predicate p2) {
3066 return (ICmpInst::isSignedPredicate(p1) == ICmpInst::isSignedPredicate(p2)) ||
3067 (ICmpInst::isSignedPredicate(p1) &&
3068 (p2 == ICmpInst::ICMP_EQ || p2 == ICmpInst::ICMP_NE)) ||
3069 (ICmpInst::isSignedPredicate(p2) &&
3070 (p1 == ICmpInst::ICMP_EQ || p1 == ICmpInst::ICMP_NE));
3074 // FoldICmpLogical - Implements (icmp1 A, B) & (icmp2 A, B) --> (icmp3 A, B)
3075 struct FoldICmpLogical {
3078 ICmpInst::Predicate pred;
3079 FoldICmpLogical(InstCombiner &ic, ICmpInst *ICI)
3080 : IC(ic), LHS(ICI->getOperand(0)), RHS(ICI->getOperand(1)),
3081 pred(ICI->getPredicate()) {}
3082 bool shouldApply(Value *V) const {
3083 if (ICmpInst *ICI = dyn_cast<ICmpInst>(V))
3084 if (PredicatesFoldable(pred, ICI->getPredicate()))
3085 return (ICI->getOperand(0) == LHS && ICI->getOperand(1) == RHS ||
3086 ICI->getOperand(0) == RHS && ICI->getOperand(1) == LHS);
3089 Instruction *apply(Instruction &Log) const {
3090 ICmpInst *ICI = cast<ICmpInst>(Log.getOperand(0));
3091 if (ICI->getOperand(0) != LHS) {
3092 assert(ICI->getOperand(1) == LHS);
3093 ICI->swapOperands(); // Swap the LHS and RHS of the ICmp
3096 ICmpInst *RHSICI = cast<ICmpInst>(Log.getOperand(1));
3097 unsigned LHSCode = getICmpCode(ICI);
3098 unsigned RHSCode = getICmpCode(RHSICI);
3100 switch (Log.getOpcode()) {
3101 case Instruction::And: Code = LHSCode & RHSCode; break;
3102 case Instruction::Or: Code = LHSCode | RHSCode; break;
3103 case Instruction::Xor: Code = LHSCode ^ RHSCode; break;
3104 default: assert(0 && "Illegal logical opcode!"); return 0;
3107 bool isSigned = ICmpInst::isSignedPredicate(RHSICI->getPredicate()) ||
3108 ICmpInst::isSignedPredicate(ICI->getPredicate());
3110 Value *RV = getICmpValue(isSigned, Code, LHS, RHS);
3111 if (Instruction *I = dyn_cast<Instruction>(RV))
3113 // Otherwise, it's a constant boolean value...
3114 return IC.ReplaceInstUsesWith(Log, RV);
3117 } // end anonymous namespace
3119 // OptAndOp - This handles expressions of the form ((val OP C1) & C2). Where
3120 // the Op parameter is 'OP', OpRHS is 'C1', and AndRHS is 'C2'. Op is
3121 // guaranteed to be a binary operator.
3122 Instruction *InstCombiner::OptAndOp(Instruction *Op,
3124 ConstantInt *AndRHS,
3125 BinaryOperator &TheAnd) {
3126 Value *X = Op->getOperand(0);
3127 Constant *Together = 0;
3129 Together = And(AndRHS, OpRHS);
3131 switch (Op->getOpcode()) {
3132 case Instruction::Xor:
3133 if (Op->hasOneUse()) {
3134 // (X ^ C1) & C2 --> (X & C2) ^ (C1&C2)
3135 Instruction *And = BinaryOperator::createAnd(X, AndRHS);
3136 InsertNewInstBefore(And, TheAnd);
3138 return BinaryOperator::createXor(And, Together);
3141 case Instruction::Or:
3142 if (Together == AndRHS) // (X | C) & C --> C
3143 return ReplaceInstUsesWith(TheAnd, AndRHS);
3145 if (Op->hasOneUse() && Together != OpRHS) {
3146 // (X | C1) & C2 --> (X | (C1&C2)) & C2
3147 Instruction *Or = BinaryOperator::createOr(X, Together);
3148 InsertNewInstBefore(Or, TheAnd);
3150 return BinaryOperator::createAnd(Or, AndRHS);
3153 case Instruction::Add:
3154 if (Op->hasOneUse()) {
3155 // Adding a one to a single bit bit-field should be turned into an XOR
3156 // of the bit. First thing to check is to see if this AND is with a
3157 // single bit constant.
3158 const APInt& AndRHSV = cast<ConstantInt>(AndRHS)->getValue();
3160 // If there is only one bit set...
3161 if (isOneBitSet(cast<ConstantInt>(AndRHS))) {
3162 // Ok, at this point, we know that we are masking the result of the
3163 // ADD down to exactly one bit. If the constant we are adding has
3164 // no bits set below this bit, then we can eliminate the ADD.
3165 const APInt& AddRHS = cast<ConstantInt>(OpRHS)->getValue();
3167 // Check to see if any bits below the one bit set in AndRHSV are set.
3168 if ((AddRHS & (AndRHSV-1)) == 0) {
3169 // If not, the only thing that can effect the output of the AND is
3170 // the bit specified by AndRHSV. If that bit is set, the effect of
3171 // the XOR is to toggle the bit. If it is clear, then the ADD has
3173 if ((AddRHS & AndRHSV) == 0) { // Bit is not set, noop
3174 TheAnd.setOperand(0, X);
3177 // Pull the XOR out of the AND.
3178 Instruction *NewAnd = BinaryOperator::createAnd(X, AndRHS);
3179 InsertNewInstBefore(NewAnd, TheAnd);
3180 NewAnd->takeName(Op);
3181 return BinaryOperator::createXor(NewAnd, AndRHS);
3188 case Instruction::Shl: {
3189 // We know that the AND will not produce any of the bits shifted in, so if
3190 // the anded constant includes them, clear them now!
3192 uint32_t BitWidth = AndRHS->getType()->getBitWidth();
3193 uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
3194 APInt ShlMask(APInt::getHighBitsSet(BitWidth, BitWidth-OpRHSVal));
3195 ConstantInt *CI = ConstantInt::get(AndRHS->getValue() & ShlMask);
3197 if (CI->getValue() == ShlMask) {
3198 // Masking out bits that the shift already masks
3199 return ReplaceInstUsesWith(TheAnd, Op); // No need for the and.
3200 } else if (CI != AndRHS) { // Reducing bits set in and.
3201 TheAnd.setOperand(1, CI);
3206 case Instruction::LShr:
3208 // We know that the AND will not produce any of the bits shifted in, so if
3209 // the anded constant includes them, clear them now! This only applies to
3210 // unsigned shifts, because a signed shr may bring in set bits!
3212 uint32_t BitWidth = AndRHS->getType()->getBitWidth();
3213 uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
3214 APInt ShrMask(APInt::getLowBitsSet(BitWidth, BitWidth - OpRHSVal));
3215 ConstantInt *CI = ConstantInt::get(AndRHS->getValue() & ShrMask);
3217 if (CI->getValue() == ShrMask) {
3218 // Masking out bits that the shift already masks.
3219 return ReplaceInstUsesWith(TheAnd, Op);
3220 } else if (CI != AndRHS) {
3221 TheAnd.setOperand(1, CI); // Reduce bits set in and cst.
3226 case Instruction::AShr:
3228 // See if this is shifting in some sign extension, then masking it out
3230 if (Op->hasOneUse()) {
3231 uint32_t BitWidth = AndRHS->getType()->getBitWidth();
3232 uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
3233 APInt ShrMask(APInt::getLowBitsSet(BitWidth, BitWidth - OpRHSVal));
3234 Constant *C = ConstantInt::get(AndRHS->getValue() & ShrMask);
3235 if (C == AndRHS) { // Masking out bits shifted in.
3236 // (Val ashr C1) & C2 -> (Val lshr C1) & C2
3237 // Make the argument unsigned.
3238 Value *ShVal = Op->getOperand(0);
3239 ShVal = InsertNewInstBefore(
3240 BinaryOperator::createLShr(ShVal, OpRHS,
3241 Op->getName()), TheAnd);
3242 return BinaryOperator::createAnd(ShVal, AndRHS, TheAnd.getName());
3251 /// InsertRangeTest - Emit a computation of: (V >= Lo && V < Hi) if Inside is
3252 /// true, otherwise (V < Lo || V >= Hi). In pratice, we emit the more efficient
3253 /// (V-Lo) <u Hi-Lo. This method expects that Lo <= Hi. isSigned indicates
3254 /// whether to treat the V, Lo and HI as signed or not. IB is the location to
3255 /// insert new instructions.
3256 Instruction *InstCombiner::InsertRangeTest(Value *V, Constant *Lo, Constant *Hi,
3257 bool isSigned, bool Inside,
3259 assert(cast<ConstantInt>(ConstantExpr::getICmp((isSigned ?
3260 ICmpInst::ICMP_SLE:ICmpInst::ICMP_ULE), Lo, Hi))->getZExtValue() &&
3261 "Lo is not <= Hi in range emission code!");
3264 if (Lo == Hi) // Trivially false.
3265 return new ICmpInst(ICmpInst::ICMP_NE, V, V);
3267 // V >= Min && V < Hi --> V < Hi
3268 if (cast<ConstantInt>(Lo)->isMinValue(isSigned)) {
3269 ICmpInst::Predicate pred = (isSigned ?
3270 ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT);
3271 return new ICmpInst(pred, V, Hi);
3274 // Emit V-Lo <u Hi-Lo
3275 Constant *NegLo = ConstantExpr::getNeg(Lo);
3276 Instruction *Add = BinaryOperator::createAdd(V, NegLo, V->getName()+".off");
3277 InsertNewInstBefore(Add, IB);
3278 Constant *UpperBound = ConstantExpr::getAdd(NegLo, Hi);
3279 return new ICmpInst(ICmpInst::ICMP_ULT, Add, UpperBound);
3282 if (Lo == Hi) // Trivially true.
3283 return new ICmpInst(ICmpInst::ICMP_EQ, V, V);
3285 // V < Min || V >= Hi -> V > Hi-1
3286 Hi = SubOne(cast<ConstantInt>(Hi));
3287 if (cast<ConstantInt>(Lo)->isMinValue(isSigned)) {
3288 ICmpInst::Predicate pred = (isSigned ?
3289 ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT);
3290 return new ICmpInst(pred, V, Hi);
3293 // Emit V-Lo >u Hi-1-Lo
3294 // Note that Hi has already had one subtracted from it, above.
3295 ConstantInt *NegLo = cast<ConstantInt>(ConstantExpr::getNeg(Lo));
3296 Instruction *Add = BinaryOperator::createAdd(V, NegLo, V->getName()+".off");
3297 InsertNewInstBefore(Add, IB);
3298 Constant *LowerBound = ConstantExpr::getAdd(NegLo, Hi);
3299 return new ICmpInst(ICmpInst::ICMP_UGT, Add, LowerBound);
3302 // isRunOfOnes - Returns true iff Val consists of one contiguous run of 1s with
3303 // any number of 0s on either side. The 1s are allowed to wrap from LSB to
3304 // MSB, so 0x000FFF0, 0x0000FFFF, and 0xFF0000FF are all runs. 0x0F0F0000 is
3305 // not, since all 1s are not contiguous.
3306 static bool isRunOfOnes(ConstantInt *Val, uint32_t &MB, uint32_t &ME) {
3307 const APInt& V = Val->getValue();
3308 uint32_t BitWidth = Val->getType()->getBitWidth();
3309 if (!APIntOps::isShiftedMask(BitWidth, V)) return false;
3311 // look for the first zero bit after the run of ones
3312 MB = BitWidth - ((V - 1) ^ V).countLeadingZeros();
3313 // look for the first non-zero bit
3314 ME = V.getActiveBits();
3318 /// FoldLogicalPlusAnd - This is part of an expression (LHS +/- RHS) & Mask,
3319 /// where isSub determines whether the operator is a sub. If we can fold one of
3320 /// the following xforms:
3322 /// ((A & N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == Mask
3323 /// ((A | N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == 0
3324 /// ((A ^ N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == 0
3326 /// return (A +/- B).
3328 Value *InstCombiner::FoldLogicalPlusAnd(Value *LHS, Value *RHS,
3329 ConstantInt *Mask, bool isSub,
3331 Instruction *LHSI = dyn_cast<Instruction>(LHS);
3332 if (!LHSI || LHSI->getNumOperands() != 2 ||
3333 !isa<ConstantInt>(LHSI->getOperand(1))) return 0;
3335 ConstantInt *N = cast<ConstantInt>(LHSI->getOperand(1));
3337 switch (LHSI->getOpcode()) {
3339 case Instruction::And:
3340 if (And(N, Mask) == Mask) {
3341 // If the AndRHS is a power of two minus one (0+1+), this is simple.
3342 if ((Mask->getValue().countLeadingZeros() +
3343 Mask->getValue().countPopulation()) ==
3344 Mask->getValue().getBitWidth())
3347 // Otherwise, if Mask is 0+1+0+, and if B is known to have the low 0+
3348 // part, we don't need any explicit masks to take them out of A. If that
3349 // is all N is, ignore it.
3350 uint32_t MB = 0, ME = 0;
3351 if (isRunOfOnes(Mask, MB, ME)) { // begin/end bit of run, inclusive
3352 uint32_t BitWidth = cast<IntegerType>(RHS->getType())->getBitWidth();
3353 APInt Mask(APInt::getLowBitsSet(BitWidth, MB-1));
3354 if (MaskedValueIsZero(RHS, Mask))
3359 case Instruction::Or:
3360 case Instruction::Xor:
3361 // If the AndRHS is a power of two minus one (0+1+), and N&Mask == 0
3362 if ((Mask->getValue().countLeadingZeros() +
3363 Mask->getValue().countPopulation()) == Mask->getValue().getBitWidth()
3364 && And(N, Mask)->isZero())
3371 New = BinaryOperator::createSub(LHSI->getOperand(0), RHS, "fold");
3373 New = BinaryOperator::createAdd(LHSI->getOperand(0), RHS, "fold");
3374 return InsertNewInstBefore(New, I);
3377 Instruction *InstCombiner::visitAnd(BinaryOperator &I) {
3378 bool Changed = SimplifyCommutative(I);
3379 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3381 if (isa<UndefValue>(Op1)) // X & undef -> 0
3382 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3386 return ReplaceInstUsesWith(I, Op1);
3388 // See if we can simplify any instructions used by the instruction whose sole
3389 // purpose is to compute bits we don't care about.
3390 if (!isa<VectorType>(I.getType())) {
3391 uint32_t BitWidth = cast<IntegerType>(I.getType())->getBitWidth();
3392 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
3393 if (SimplifyDemandedBits(&I, APInt::getAllOnesValue(BitWidth),
3394 KnownZero, KnownOne))
3397 if (ConstantVector *CP = dyn_cast<ConstantVector>(Op1)) {
3398 if (CP->isAllOnesValue()) // X & <-1,-1> -> X
3399 return ReplaceInstUsesWith(I, I.getOperand(0));
3400 } else if (isa<ConstantAggregateZero>(Op1)) {
3401 return ReplaceInstUsesWith(I, Op1); // X & <0,0> -> <0,0>
3405 if (ConstantInt *AndRHS = dyn_cast<ConstantInt>(Op1)) {
3406 const APInt& AndRHSMask = AndRHS->getValue();
3407 APInt NotAndRHS(~AndRHSMask);
3409 // Optimize a variety of ((val OP C1) & C2) combinations...
3410 if (isa<BinaryOperator>(Op0)) {
3411 Instruction *Op0I = cast<Instruction>(Op0);
3412 Value *Op0LHS = Op0I->getOperand(0);
3413 Value *Op0RHS = Op0I->getOperand(1);
3414 switch (Op0I->getOpcode()) {
3415 case Instruction::Xor:
3416 case Instruction::Or:
3417 // If the mask is only needed on one incoming arm, push it up.
3418 if (Op0I->hasOneUse()) {
3419 if (MaskedValueIsZero(Op0LHS, NotAndRHS)) {
3420 // Not masking anything out for the LHS, move to RHS.
3421 Instruction *NewRHS = BinaryOperator::createAnd(Op0RHS, AndRHS,
3422 Op0RHS->getName()+".masked");
3423 InsertNewInstBefore(NewRHS, I);
3424 return BinaryOperator::create(
3425 cast<BinaryOperator>(Op0I)->getOpcode(), Op0LHS, NewRHS);
3427 if (!isa<Constant>(Op0RHS) &&
3428 MaskedValueIsZero(Op0RHS, NotAndRHS)) {
3429 // Not masking anything out for the RHS, move to LHS.
3430 Instruction *NewLHS = BinaryOperator::createAnd(Op0LHS, AndRHS,
3431 Op0LHS->getName()+".masked");
3432 InsertNewInstBefore(NewLHS, I);
3433 return BinaryOperator::create(
3434 cast<BinaryOperator>(Op0I)->getOpcode(), NewLHS, Op0RHS);
3439 case Instruction::Add:
3440 // ((A & N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == AndRHS.
3441 // ((A | N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == 0
3442 // ((A ^ N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == 0
3443 if (Value *V = FoldLogicalPlusAnd(Op0LHS, Op0RHS, AndRHS, false, I))
3444 return BinaryOperator::createAnd(V, AndRHS);
3445 if (Value *V = FoldLogicalPlusAnd(Op0RHS, Op0LHS, AndRHS, false, I))
3446 return BinaryOperator::createAnd(V, AndRHS); // Add commutes
3449 case Instruction::Sub:
3450 // ((A & N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == AndRHS.
3451 // ((A | N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == 0
3452 // ((A ^ N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == 0
3453 if (Value *V = FoldLogicalPlusAnd(Op0LHS, Op0RHS, AndRHS, true, I))
3454 return BinaryOperator::createAnd(V, AndRHS);
3458 if (ConstantInt *Op0CI = dyn_cast<ConstantInt>(Op0I->getOperand(1)))
3459 if (Instruction *Res = OptAndOp(Op0I, Op0CI, AndRHS, I))
3461 } else if (CastInst *CI = dyn_cast<CastInst>(Op0)) {
3462 // If this is an integer truncation or change from signed-to-unsigned, and
3463 // if the source is an and/or with immediate, transform it. This
3464 // frequently occurs for bitfield accesses.
3465 if (Instruction *CastOp = dyn_cast<Instruction>(CI->getOperand(0))) {
3466 if ((isa<TruncInst>(CI) || isa<BitCastInst>(CI)) &&
3467 CastOp->getNumOperands() == 2)
3468 if (ConstantInt *AndCI = dyn_cast<ConstantInt>(CastOp->getOperand(1)))
3469 if (CastOp->getOpcode() == Instruction::And) {
3470 // Change: and (cast (and X, C1) to T), C2
3471 // into : and (cast X to T), trunc_or_bitcast(C1)&C2
3472 // This will fold the two constants together, which may allow
3473 // other simplifications.
3474 Instruction *NewCast = CastInst::createTruncOrBitCast(
3475 CastOp->getOperand(0), I.getType(),
3476 CastOp->getName()+".shrunk");
3477 NewCast = InsertNewInstBefore(NewCast, I);
3478 // trunc_or_bitcast(C1)&C2
3479 Constant *C3 = ConstantExpr::getTruncOrBitCast(AndCI,I.getType());
3480 C3 = ConstantExpr::getAnd(C3, AndRHS);
3481 return BinaryOperator::createAnd(NewCast, C3);
3482 } else if (CastOp->getOpcode() == Instruction::Or) {
3483 // Change: and (cast (or X, C1) to T), C2
3484 // into : trunc(C1)&C2 iff trunc(C1)&C2 == C2
3485 Constant *C3 = ConstantExpr::getTruncOrBitCast(AndCI,I.getType());
3486 if (ConstantExpr::getAnd(C3, AndRHS) == AndRHS) // trunc(C1)&C2
3487 return ReplaceInstUsesWith(I, AndRHS);
3492 // Try to fold constant and into select arguments.
3493 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
3494 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
3496 if (isa<PHINode>(Op0))
3497 if (Instruction *NV = FoldOpIntoPhi(I))
3501 Value *Op0NotVal = dyn_castNotVal(Op0);
3502 Value *Op1NotVal = dyn_castNotVal(Op1);
3504 if (Op0NotVal == Op1 || Op1NotVal == Op0) // A & ~A == ~A & A == 0
3505 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3507 // (~A & ~B) == (~(A | B)) - De Morgan's Law
3508 if (Op0NotVal && Op1NotVal && isOnlyUse(Op0) && isOnlyUse(Op1)) {
3509 Instruction *Or = BinaryOperator::createOr(Op0NotVal, Op1NotVal,
3510 I.getName()+".demorgan");
3511 InsertNewInstBefore(Or, I);
3512 return BinaryOperator::createNot(Or);
3516 Value *A = 0, *B = 0, *C = 0, *D = 0;
3517 if (match(Op0, m_Or(m_Value(A), m_Value(B)))) {
3518 if (A == Op1 || B == Op1) // (A | ?) & A --> A
3519 return ReplaceInstUsesWith(I, Op1);
3521 // (A|B) & ~(A&B) -> A^B
3522 if (match(Op1, m_Not(m_And(m_Value(C), m_Value(D))))) {
3523 if ((A == C && B == D) || (A == D && B == C))
3524 return BinaryOperator::createXor(A, B);
3528 if (match(Op1, m_Or(m_Value(A), m_Value(B)))) {
3529 if (A == Op0 || B == Op0) // A & (A | ?) --> A
3530 return ReplaceInstUsesWith(I, Op0);
3532 // ~(A&B) & (A|B) -> A^B
3533 if (match(Op0, m_Not(m_And(m_Value(C), m_Value(D))))) {
3534 if ((A == C && B == D) || (A == D && B == C))
3535 return BinaryOperator::createXor(A, B);
3539 if (Op0->hasOneUse() &&
3540 match(Op0, m_Xor(m_Value(A), m_Value(B)))) {
3541 if (A == Op1) { // (A^B)&A -> A&(A^B)
3542 I.swapOperands(); // Simplify below
3543 std::swap(Op0, Op1);
3544 } else if (B == Op1) { // (A^B)&B -> B&(B^A)
3545 cast<BinaryOperator>(Op0)->swapOperands();
3546 I.swapOperands(); // Simplify below
3547 std::swap(Op0, Op1);
3550 if (Op1->hasOneUse() &&
3551 match(Op1, m_Xor(m_Value(A), m_Value(B)))) {
3552 if (B == Op0) { // B&(A^B) -> B&(B^A)
3553 cast<BinaryOperator>(Op1)->swapOperands();
3556 if (A == Op0) { // A&(A^B) -> A & ~B
3557 Instruction *NotB = BinaryOperator::createNot(B, "tmp");
3558 InsertNewInstBefore(NotB, I);
3559 return BinaryOperator::createAnd(A, NotB);
3564 if (ICmpInst *RHS = dyn_cast<ICmpInst>(Op1)) {
3565 // (icmp1 A, B) & (icmp2 A, B) --> (icmp3 A, B)
3566 if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS)))
3569 Value *LHSVal, *RHSVal;
3570 ConstantInt *LHSCst, *RHSCst;
3571 ICmpInst::Predicate LHSCC, RHSCC;
3572 if (match(Op0, m_ICmp(LHSCC, m_Value(LHSVal), m_ConstantInt(LHSCst))))
3573 if (match(RHS, m_ICmp(RHSCC, m_Value(RHSVal), m_ConstantInt(RHSCst))))
3574 if (LHSVal == RHSVal && // Found (X icmp C1) & (X icmp C2)
3575 // ICMP_[GL]E X, CST is folded to ICMP_[GL]T elsewhere.
3576 LHSCC != ICmpInst::ICMP_UGE && LHSCC != ICmpInst::ICMP_ULE &&
3577 RHSCC != ICmpInst::ICMP_UGE && RHSCC != ICmpInst::ICMP_ULE &&
3578 LHSCC != ICmpInst::ICMP_SGE && LHSCC != ICmpInst::ICMP_SLE &&
3579 RHSCC != ICmpInst::ICMP_SGE && RHSCC != ICmpInst::ICMP_SLE &&
3581 // Don't try to fold ICMP_SLT + ICMP_ULT.
3582 (ICmpInst::isEquality(LHSCC) || ICmpInst::isEquality(RHSCC) ||
3583 ICmpInst::isSignedPredicate(LHSCC) ==
3584 ICmpInst::isSignedPredicate(RHSCC))) {
3585 // Ensure that the larger constant is on the RHS.
3586 ICmpInst::Predicate GT;
3587 if (ICmpInst::isSignedPredicate(LHSCC) ||
3588 (ICmpInst::isEquality(LHSCC) &&
3589 ICmpInst::isSignedPredicate(RHSCC)))
3590 GT = ICmpInst::ICMP_SGT;
3592 GT = ICmpInst::ICMP_UGT;
3594 Constant *Cmp = ConstantExpr::getICmp(GT, LHSCst, RHSCst);
3595 ICmpInst *LHS = cast<ICmpInst>(Op0);
3596 if (cast<ConstantInt>(Cmp)->getZExtValue()) {
3597 std::swap(LHS, RHS);
3598 std::swap(LHSCst, RHSCst);
3599 std::swap(LHSCC, RHSCC);
3602 // At this point, we know we have have two icmp instructions
3603 // comparing a value against two constants and and'ing the result
3604 // together. Because of the above check, we know that we only have
3605 // icmp eq, icmp ne, icmp [su]lt, and icmp [SU]gt here. We also know
3606 // (from the FoldICmpLogical check above), that the two constants
3607 // are not equal and that the larger constant is on the RHS
3608 assert(LHSCst != RHSCst && "Compares not folded above?");
3611 default: assert(0 && "Unknown integer condition code!");
3612 case ICmpInst::ICMP_EQ:
3614 default: assert(0 && "Unknown integer condition code!");
3615 case ICmpInst::ICMP_EQ: // (X == 13 & X == 15) -> false
3616 case ICmpInst::ICMP_UGT: // (X == 13 & X > 15) -> false
3617 case ICmpInst::ICMP_SGT: // (X == 13 & X > 15) -> false
3618 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
3619 case ICmpInst::ICMP_NE: // (X == 13 & X != 15) -> X == 13
3620 case ICmpInst::ICMP_ULT: // (X == 13 & X < 15) -> X == 13
3621 case ICmpInst::ICMP_SLT: // (X == 13 & X < 15) -> X == 13
3622 return ReplaceInstUsesWith(I, LHS);
3624 case ICmpInst::ICMP_NE:
3626 default: assert(0 && "Unknown integer condition code!");
3627 case ICmpInst::ICMP_ULT:
3628 if (LHSCst == SubOne(RHSCst)) // (X != 13 & X u< 14) -> X < 13
3629 return new ICmpInst(ICmpInst::ICMP_ULT, LHSVal, LHSCst);
3630 break; // (X != 13 & X u< 15) -> no change
3631 case ICmpInst::ICMP_SLT:
3632 if (LHSCst == SubOne(RHSCst)) // (X != 13 & X s< 14) -> X < 13
3633 return new ICmpInst(ICmpInst::ICMP_SLT, LHSVal, LHSCst);
3634 break; // (X != 13 & X s< 15) -> no change
3635 case ICmpInst::ICMP_EQ: // (X != 13 & X == 15) -> X == 15
3636 case ICmpInst::ICMP_UGT: // (X != 13 & X u> 15) -> X u> 15
3637 case ICmpInst::ICMP_SGT: // (X != 13 & X s> 15) -> X s> 15
3638 return ReplaceInstUsesWith(I, RHS);
3639 case ICmpInst::ICMP_NE:
3640 if (LHSCst == SubOne(RHSCst)){// (X != 13 & X != 14) -> X-13 >u 1
3641 Constant *AddCST = ConstantExpr::getNeg(LHSCst);
3642 Instruction *Add = BinaryOperator::createAdd(LHSVal, AddCST,
3643 LHSVal->getName()+".off");
3644 InsertNewInstBefore(Add, I);
3645 return new ICmpInst(ICmpInst::ICMP_UGT, Add,
3646 ConstantInt::get(Add->getType(), 1));
3648 break; // (X != 13 & X != 15) -> no change
3651 case ICmpInst::ICMP_ULT:
3653 default: assert(0 && "Unknown integer condition code!");
3654 case ICmpInst::ICMP_EQ: // (X u< 13 & X == 15) -> false
3655 case ICmpInst::ICMP_UGT: // (X u< 13 & X u> 15) -> false
3656 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
3657 case ICmpInst::ICMP_SGT: // (X u< 13 & X s> 15) -> no change
3659 case ICmpInst::ICMP_NE: // (X u< 13 & X != 15) -> X u< 13
3660 case ICmpInst::ICMP_ULT: // (X u< 13 & X u< 15) -> X u< 13
3661 return ReplaceInstUsesWith(I, LHS);
3662 case ICmpInst::ICMP_SLT: // (X u< 13 & X s< 15) -> no change
3666 case ICmpInst::ICMP_SLT:
3668 default: assert(0 && "Unknown integer condition code!");
3669 case ICmpInst::ICMP_EQ: // (X s< 13 & X == 15) -> false
3670 case ICmpInst::ICMP_SGT: // (X s< 13 & X s> 15) -> false
3671 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
3672 case ICmpInst::ICMP_UGT: // (X s< 13 & X u> 15) -> no change
3674 case ICmpInst::ICMP_NE: // (X s< 13 & X != 15) -> X < 13
3675 case ICmpInst::ICMP_SLT: // (X s< 13 & X s< 15) -> X < 13
3676 return ReplaceInstUsesWith(I, LHS);
3677 case ICmpInst::ICMP_ULT: // (X s< 13 & X u< 15) -> no change
3681 case ICmpInst::ICMP_UGT:
3683 default: assert(0 && "Unknown integer condition code!");
3684 case ICmpInst::ICMP_EQ: // (X u> 13 & X == 15) -> X > 13
3685 return ReplaceInstUsesWith(I, LHS);
3686 case ICmpInst::ICMP_UGT: // (X u> 13 & X u> 15) -> X u> 15
3687 return ReplaceInstUsesWith(I, RHS);
3688 case ICmpInst::ICMP_SGT: // (X u> 13 & X s> 15) -> no change
3690 case ICmpInst::ICMP_NE:
3691 if (RHSCst == AddOne(LHSCst)) // (X u> 13 & X != 14) -> X u> 14
3692 return new ICmpInst(LHSCC, LHSVal, RHSCst);
3693 break; // (X u> 13 & X != 15) -> no change
3694 case ICmpInst::ICMP_ULT: // (X u> 13 & X u< 15) ->(X-14) <u 1
3695 return InsertRangeTest(LHSVal, AddOne(LHSCst), RHSCst, false,
3697 case ICmpInst::ICMP_SLT: // (X u> 13 & X s< 15) -> no change
3701 case ICmpInst::ICMP_SGT:
3703 default: assert(0 && "Unknown integer condition code!");
3704 case ICmpInst::ICMP_EQ: // (X s> 13 & X == 15) -> X == 15
3705 case ICmpInst::ICMP_SGT: // (X s> 13 & X s> 15) -> X s> 15
3706 return ReplaceInstUsesWith(I, RHS);
3707 case ICmpInst::ICMP_UGT: // (X s> 13 & X u> 15) -> no change
3709 case ICmpInst::ICMP_NE:
3710 if (RHSCst == AddOne(LHSCst)) // (X s> 13 & X != 14) -> X s> 14
3711 return new ICmpInst(LHSCC, LHSVal, RHSCst);
3712 break; // (X s> 13 & X != 15) -> no change
3713 case ICmpInst::ICMP_SLT: // (X s> 13 & X s< 15) ->(X-14) s< 1
3714 return InsertRangeTest(LHSVal, AddOne(LHSCst), RHSCst, true,
3716 case ICmpInst::ICMP_ULT: // (X s> 13 & X u< 15) -> no change
3724 // fold (and (cast A), (cast B)) -> (cast (and A, B))
3725 if (CastInst *Op0C = dyn_cast<CastInst>(Op0))
3726 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
3727 if (Op0C->getOpcode() == Op1C->getOpcode()) { // same cast kind ?
3728 const Type *SrcTy = Op0C->getOperand(0)->getType();
3729 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isInteger() &&
3730 // Only do this if the casts both really cause code to be generated.
3731 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
3733 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
3735 Instruction *NewOp = BinaryOperator::createAnd(Op0C->getOperand(0),
3736 Op1C->getOperand(0),
3738 InsertNewInstBefore(NewOp, I);
3739 return CastInst::create(Op0C->getOpcode(), NewOp, I.getType());
3743 // (X >> Z) & (Y >> Z) -> (X&Y) >> Z for all shifts.
3744 if (BinaryOperator *SI1 = dyn_cast<BinaryOperator>(Op1)) {
3745 if (BinaryOperator *SI0 = dyn_cast<BinaryOperator>(Op0))
3746 if (SI0->isShift() && SI0->getOpcode() == SI1->getOpcode() &&
3747 SI0->getOperand(1) == SI1->getOperand(1) &&
3748 (SI0->hasOneUse() || SI1->hasOneUse())) {
3749 Instruction *NewOp =
3750 InsertNewInstBefore(BinaryOperator::createAnd(SI0->getOperand(0),
3752 SI0->getName()), I);
3753 return BinaryOperator::create(SI1->getOpcode(), NewOp,
3754 SI1->getOperand(1));
3758 // (fcmp ord x, c) & (fcmp ord y, c) -> (fcmp ord x, y)
3759 if (FCmpInst *LHS = dyn_cast<FCmpInst>(I.getOperand(0))) {
3760 if (FCmpInst *RHS = dyn_cast<FCmpInst>(I.getOperand(1))) {
3761 if (LHS->getPredicate() == FCmpInst::FCMP_ORD &&
3762 RHS->getPredicate() == FCmpInst::FCMP_ORD)
3763 if (ConstantFP *LHSC = dyn_cast<ConstantFP>(LHS->getOperand(1)))
3764 if (ConstantFP *RHSC = dyn_cast<ConstantFP>(RHS->getOperand(1))) {
3765 // If either of the constants are nans, then the whole thing returns
3767 if (LHSC->getValueAPF().isNaN() || RHSC->getValueAPF().isNaN())
3768 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
3769 return new FCmpInst(FCmpInst::FCMP_ORD, LHS->getOperand(0),
3770 RHS->getOperand(0));
3775 return Changed ? &I : 0;
3778 /// CollectBSwapParts - Look to see if the specified value defines a single byte
3779 /// in the result. If it does, and if the specified byte hasn't been filled in
3780 /// yet, fill it in and return false.
3781 static bool CollectBSwapParts(Value *V, SmallVector<Value*, 8> &ByteValues) {
3782 Instruction *I = dyn_cast<Instruction>(V);
3783 if (I == 0) return true;
3785 // If this is an or instruction, it is an inner node of the bswap.
3786 if (I->getOpcode() == Instruction::Or)
3787 return CollectBSwapParts(I->getOperand(0), ByteValues) ||
3788 CollectBSwapParts(I->getOperand(1), ByteValues);
3790 uint32_t BitWidth = I->getType()->getPrimitiveSizeInBits();
3791 // If this is a shift by a constant int, and it is "24", then its operand
3792 // defines a byte. We only handle unsigned types here.
3793 if (I->isShift() && isa<ConstantInt>(I->getOperand(1))) {
3794 // Not shifting the entire input by N-1 bytes?
3795 if (cast<ConstantInt>(I->getOperand(1))->getLimitedValue(BitWidth) !=
3796 8*(ByteValues.size()-1))
3800 if (I->getOpcode() == Instruction::Shl) {
3801 // X << 24 defines the top byte with the lowest of the input bytes.
3802 DestNo = ByteValues.size()-1;
3804 // X >>u 24 defines the low byte with the highest of the input bytes.
3808 // If the destination byte value is already defined, the values are or'd
3809 // together, which isn't a bswap (unless it's an or of the same bits).
3810 if (ByteValues[DestNo] && ByteValues[DestNo] != I->getOperand(0))
3812 ByteValues[DestNo] = I->getOperand(0);
3816 // Otherwise, we can only handle and(shift X, imm), imm). Bail out of if we
3818 Value *Shift = 0, *ShiftLHS = 0;
3819 ConstantInt *AndAmt = 0, *ShiftAmt = 0;
3820 if (!match(I, m_And(m_Value(Shift), m_ConstantInt(AndAmt))) ||
3821 !match(Shift, m_Shift(m_Value(ShiftLHS), m_ConstantInt(ShiftAmt))))
3823 Instruction *SI = cast<Instruction>(Shift);
3825 // Make sure that the shift amount is by a multiple of 8 and isn't too big.
3826 if (ShiftAmt->getLimitedValue(BitWidth) & 7 ||
3827 ShiftAmt->getLimitedValue(BitWidth) > 8*ByteValues.size())
3830 // Turn 0xFF -> 0, 0xFF00 -> 1, 0xFF0000 -> 2, etc.
3832 if (AndAmt->getValue().getActiveBits() > 64)
3834 uint64_t AndAmtVal = AndAmt->getZExtValue();
3835 for (DestByte = 0; DestByte != ByteValues.size(); ++DestByte)
3836 if (AndAmtVal == uint64_t(0xFF) << 8*DestByte)
3838 // Unknown mask for bswap.
3839 if (DestByte == ByteValues.size()) return true;
3841 unsigned ShiftBytes = ShiftAmt->getZExtValue()/8;
3843 if (SI->getOpcode() == Instruction::Shl)
3844 SrcByte = DestByte - ShiftBytes;
3846 SrcByte = DestByte + ShiftBytes;
3848 // If the SrcByte isn't a bswapped value from the DestByte, reject it.
3849 if (SrcByte != ByteValues.size()-DestByte-1)
3852 // If the destination byte value is already defined, the values are or'd
3853 // together, which isn't a bswap (unless it's an or of the same bits).
3854 if (ByteValues[DestByte] && ByteValues[DestByte] != SI->getOperand(0))
3856 ByteValues[DestByte] = SI->getOperand(0);
3860 /// MatchBSwap - Given an OR instruction, check to see if this is a bswap idiom.
3861 /// If so, insert the new bswap intrinsic and return it.
3862 Instruction *InstCombiner::MatchBSwap(BinaryOperator &I) {
3863 const IntegerType *ITy = dyn_cast<IntegerType>(I.getType());
3864 if (!ITy || ITy->getBitWidth() % 16)
3865 return 0; // Can only bswap pairs of bytes. Can't do vectors.
3867 /// ByteValues - For each byte of the result, we keep track of which value
3868 /// defines each byte.
3869 SmallVector<Value*, 8> ByteValues;
3870 ByteValues.resize(ITy->getBitWidth()/8);
3872 // Try to find all the pieces corresponding to the bswap.
3873 if (CollectBSwapParts(I.getOperand(0), ByteValues) ||
3874 CollectBSwapParts(I.getOperand(1), ByteValues))
3877 // Check to see if all of the bytes come from the same value.
3878 Value *V = ByteValues[0];
3879 if (V == 0) return 0; // Didn't find a byte? Must be zero.
3881 // Check to make sure that all of the bytes come from the same value.
3882 for (unsigned i = 1, e = ByteValues.size(); i != e; ++i)
3883 if (ByteValues[i] != V)
3885 const Type *Tys[] = { ITy };
3886 Module *M = I.getParent()->getParent()->getParent();
3887 Function *F = Intrinsic::getDeclaration(M, Intrinsic::bswap, Tys, 1);
3888 return new CallInst(F, V);
3892 Instruction *InstCombiner::visitOr(BinaryOperator &I) {
3893 bool Changed = SimplifyCommutative(I);
3894 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3896 if (isa<UndefValue>(Op1)) // X | undef -> -1
3897 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
3901 return ReplaceInstUsesWith(I, Op0);
3903 // See if we can simplify any instructions used by the instruction whose sole
3904 // purpose is to compute bits we don't care about.
3905 if (!isa<VectorType>(I.getType())) {
3906 uint32_t BitWidth = cast<IntegerType>(I.getType())->getBitWidth();
3907 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
3908 if (SimplifyDemandedBits(&I, APInt::getAllOnesValue(BitWidth),
3909 KnownZero, KnownOne))
3911 } else if (isa<ConstantAggregateZero>(Op1)) {
3912 return ReplaceInstUsesWith(I, Op0); // X | <0,0> -> X
3913 } else if (ConstantVector *CP = dyn_cast<ConstantVector>(Op1)) {
3914 if (CP->isAllOnesValue()) // X | <-1,-1> -> <-1,-1>
3915 return ReplaceInstUsesWith(I, I.getOperand(1));
3921 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
3922 ConstantInt *C1 = 0; Value *X = 0;
3923 // (X & C1) | C2 --> (X | C2) & (C1|C2)
3924 if (match(Op0, m_And(m_Value(X), m_ConstantInt(C1))) && isOnlyUse(Op0)) {
3925 Instruction *Or = BinaryOperator::createOr(X, RHS);
3926 InsertNewInstBefore(Or, I);
3928 return BinaryOperator::createAnd(Or,
3929 ConstantInt::get(RHS->getValue() | C1->getValue()));
3932 // (X ^ C1) | C2 --> (X | C2) ^ (C1&~C2)
3933 if (match(Op0, m_Xor(m_Value(X), m_ConstantInt(C1))) && isOnlyUse(Op0)) {
3934 Instruction *Or = BinaryOperator::createOr(X, RHS);
3935 InsertNewInstBefore(Or, I);
3937 return BinaryOperator::createXor(Or,
3938 ConstantInt::get(C1->getValue() & ~RHS->getValue()));
3941 // Try to fold constant and into select arguments.
3942 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
3943 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
3945 if (isa<PHINode>(Op0))
3946 if (Instruction *NV = FoldOpIntoPhi(I))
3950 Value *A = 0, *B = 0;
3951 ConstantInt *C1 = 0, *C2 = 0;
3953 if (match(Op0, m_And(m_Value(A), m_Value(B))))
3954 if (A == Op1 || B == Op1) // (A & ?) | A --> A
3955 return ReplaceInstUsesWith(I, Op1);
3956 if (match(Op1, m_And(m_Value(A), m_Value(B))))
3957 if (A == Op0 || B == Op0) // A | (A & ?) --> A
3958 return ReplaceInstUsesWith(I, Op0);
3960 // (A | B) | C and A | (B | C) -> bswap if possible.
3961 // (A >> B) | (C << D) and (A << B) | (B >> C) -> bswap if possible.
3962 if (match(Op0, m_Or(m_Value(), m_Value())) ||
3963 match(Op1, m_Or(m_Value(), m_Value())) ||
3964 (match(Op0, m_Shift(m_Value(), m_Value())) &&
3965 match(Op1, m_Shift(m_Value(), m_Value())))) {
3966 if (Instruction *BSwap = MatchBSwap(I))
3970 // (X^C)|Y -> (X|Y)^C iff Y&C == 0
3971 if (Op0->hasOneUse() && match(Op0, m_Xor(m_Value(A), m_ConstantInt(C1))) &&
3972 MaskedValueIsZero(Op1, C1->getValue())) {
3973 Instruction *NOr = BinaryOperator::createOr(A, Op1);
3974 InsertNewInstBefore(NOr, I);
3976 return BinaryOperator::createXor(NOr, C1);
3979 // Y|(X^C) -> (X|Y)^C iff Y&C == 0
3980 if (Op1->hasOneUse() && match(Op1, m_Xor(m_Value(A), m_ConstantInt(C1))) &&
3981 MaskedValueIsZero(Op0, C1->getValue())) {
3982 Instruction *NOr = BinaryOperator::createOr(A, Op0);
3983 InsertNewInstBefore(NOr, I);
3985 return BinaryOperator::createXor(NOr, C1);
3989 Value *C = 0, *D = 0;
3990 if (match(Op0, m_And(m_Value(A), m_Value(C))) &&
3991 match(Op1, m_And(m_Value(B), m_Value(D)))) {
3992 Value *V1 = 0, *V2 = 0, *V3 = 0;
3993 C1 = dyn_cast<ConstantInt>(C);
3994 C2 = dyn_cast<ConstantInt>(D);
3995 if (C1 && C2) { // (A & C1)|(B & C2)
3996 // If we have: ((V + N) & C1) | (V & C2)
3997 // .. and C2 = ~C1 and C2 is 0+1+ and (N & C2) == 0
3998 // replace with V+N.
3999 if (C1->getValue() == ~C2->getValue()) {
4000 if ((C2->getValue() & (C2->getValue()+1)) == 0 && // C2 == 0+1+
4001 match(A, m_Add(m_Value(V1), m_Value(V2)))) {
4002 // Add commutes, try both ways.
4003 if (V1 == B && MaskedValueIsZero(V2, C2->getValue()))
4004 return ReplaceInstUsesWith(I, A);
4005 if (V2 == B && MaskedValueIsZero(V1, C2->getValue()))
4006 return ReplaceInstUsesWith(I, A);
4008 // Or commutes, try both ways.
4009 if ((C1->getValue() & (C1->getValue()+1)) == 0 &&
4010 match(B, m_Add(m_Value(V1), m_Value(V2)))) {
4011 // Add commutes, try both ways.
4012 if (V1 == A && MaskedValueIsZero(V2, C1->getValue()))
4013 return ReplaceInstUsesWith(I, B);
4014 if (V2 == A && MaskedValueIsZero(V1, C1->getValue()))
4015 return ReplaceInstUsesWith(I, B);
4018 V1 = 0; V2 = 0; V3 = 0;
4021 // Check to see if we have any common things being and'ed. If so, find the
4022 // terms for V1 & (V2|V3).
4023 if (isOnlyUse(Op0) || isOnlyUse(Op1)) {
4024 if (A == B) // (A & C)|(A & D) == A & (C|D)
4025 V1 = A, V2 = C, V3 = D;
4026 else if (A == D) // (A & C)|(B & A) == A & (B|C)
4027 V1 = A, V2 = B, V3 = C;
4028 else if (C == B) // (A & C)|(C & D) == C & (A|D)
4029 V1 = C, V2 = A, V3 = D;
4030 else if (C == D) // (A & C)|(B & C) == C & (A|B)
4031 V1 = C, V2 = A, V3 = B;
4035 InsertNewInstBefore(BinaryOperator::createOr(V2, V3, "tmp"), I);
4036 return BinaryOperator::createAnd(V1, Or);
4041 // (X >> Z) | (Y >> Z) -> (X|Y) >> Z for all shifts.
4042 if (BinaryOperator *SI1 = dyn_cast<BinaryOperator>(Op1)) {
4043 if (BinaryOperator *SI0 = dyn_cast<BinaryOperator>(Op0))
4044 if (SI0->isShift() && SI0->getOpcode() == SI1->getOpcode() &&
4045 SI0->getOperand(1) == SI1->getOperand(1) &&
4046 (SI0->hasOneUse() || SI1->hasOneUse())) {
4047 Instruction *NewOp =
4048 InsertNewInstBefore(BinaryOperator::createOr(SI0->getOperand(0),
4050 SI0->getName()), I);
4051 return BinaryOperator::create(SI1->getOpcode(), NewOp,
4052 SI1->getOperand(1));
4056 if (match(Op0, m_Not(m_Value(A)))) { // ~A | Op1
4057 if (A == Op1) // ~A | A == -1
4058 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
4062 // Note, A is still live here!
4063 if (match(Op1, m_Not(m_Value(B)))) { // Op0 | ~B
4065 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
4067 // (~A | ~B) == (~(A & B)) - De Morgan's Law
4068 if (A && isOnlyUse(Op0) && isOnlyUse(Op1)) {
4069 Value *And = InsertNewInstBefore(BinaryOperator::createAnd(A, B,
4070 I.getName()+".demorgan"), I);
4071 return BinaryOperator::createNot(And);
4075 // (icmp1 A, B) | (icmp2 A, B) --> (icmp3 A, B)
4076 if (ICmpInst *RHS = dyn_cast<ICmpInst>(I.getOperand(1))) {
4077 if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS)))
4080 Value *LHSVal, *RHSVal;
4081 ConstantInt *LHSCst, *RHSCst;
4082 ICmpInst::Predicate LHSCC, RHSCC;
4083 if (match(Op0, m_ICmp(LHSCC, m_Value(LHSVal), m_ConstantInt(LHSCst))))
4084 if (match(RHS, m_ICmp(RHSCC, m_Value(RHSVal), m_ConstantInt(RHSCst))))
4085 if (LHSVal == RHSVal && // Found (X icmp C1) | (X icmp C2)
4086 // icmp [us][gl]e x, cst is folded to icmp [us][gl]t elsewhere.
4087 LHSCC != ICmpInst::ICMP_UGE && LHSCC != ICmpInst::ICMP_ULE &&
4088 RHSCC != ICmpInst::ICMP_UGE && RHSCC != ICmpInst::ICMP_ULE &&
4089 LHSCC != ICmpInst::ICMP_SGE && LHSCC != ICmpInst::ICMP_SLE &&
4090 RHSCC != ICmpInst::ICMP_SGE && RHSCC != ICmpInst::ICMP_SLE &&
4091 // We can't fold (ugt x, C) | (sgt x, C2).
4092 PredicatesFoldable(LHSCC, RHSCC)) {
4093 // Ensure that the larger constant is on the RHS.
4094 ICmpInst *LHS = cast<ICmpInst>(Op0);
4096 if (ICmpInst::isSignedPredicate(LHSCC))
4097 NeedsSwap = LHSCst->getValue().sgt(RHSCst->getValue());
4099 NeedsSwap = LHSCst->getValue().ugt(RHSCst->getValue());
4102 std::swap(LHS, RHS);
4103 std::swap(LHSCst, RHSCst);
4104 std::swap(LHSCC, RHSCC);
4107 // At this point, we know we have have two icmp instructions
4108 // comparing a value against two constants and or'ing the result
4109 // together. Because of the above check, we know that we only have
4110 // ICMP_EQ, ICMP_NE, ICMP_LT, and ICMP_GT here. We also know (from the
4111 // FoldICmpLogical check above), that the two constants are not
4113 assert(LHSCst != RHSCst && "Compares not folded above?");
4116 default: assert(0 && "Unknown integer condition code!");
4117 case ICmpInst::ICMP_EQ:
4119 default: assert(0 && "Unknown integer condition code!");
4120 case ICmpInst::ICMP_EQ:
4121 if (LHSCst == SubOne(RHSCst)) {// (X == 13 | X == 14) -> X-13 <u 2
4122 Constant *AddCST = ConstantExpr::getNeg(LHSCst);
4123 Instruction *Add = BinaryOperator::createAdd(LHSVal, AddCST,
4124 LHSVal->getName()+".off");
4125 InsertNewInstBefore(Add, I);
4126 AddCST = Subtract(AddOne(RHSCst), LHSCst);
4127 return new ICmpInst(ICmpInst::ICMP_ULT, Add, AddCST);
4129 break; // (X == 13 | X == 15) -> no change
4130 case ICmpInst::ICMP_UGT: // (X == 13 | X u> 14) -> no change
4131 case ICmpInst::ICMP_SGT: // (X == 13 | X s> 14) -> no change
4133 case ICmpInst::ICMP_NE: // (X == 13 | X != 15) -> X != 15
4134 case ICmpInst::ICMP_ULT: // (X == 13 | X u< 15) -> X u< 15
4135 case ICmpInst::ICMP_SLT: // (X == 13 | X s< 15) -> X s< 15
4136 return ReplaceInstUsesWith(I, RHS);
4139 case ICmpInst::ICMP_NE:
4141 default: assert(0 && "Unknown integer condition code!");
4142 case ICmpInst::ICMP_EQ: // (X != 13 | X == 15) -> X != 13
4143 case ICmpInst::ICMP_UGT: // (X != 13 | X u> 15) -> X != 13
4144 case ICmpInst::ICMP_SGT: // (X != 13 | X s> 15) -> X != 13
4145 return ReplaceInstUsesWith(I, LHS);
4146 case ICmpInst::ICMP_NE: // (X != 13 | X != 15) -> true
4147 case ICmpInst::ICMP_ULT: // (X != 13 | X u< 15) -> true
4148 case ICmpInst::ICMP_SLT: // (X != 13 | X s< 15) -> true
4149 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4152 case ICmpInst::ICMP_ULT:
4154 default: assert(0 && "Unknown integer condition code!");
4155 case ICmpInst::ICMP_EQ: // (X u< 13 | X == 14) -> no change
4157 case ICmpInst::ICMP_UGT: // (X u< 13 | X u> 15) ->(X-13) u> 2
4158 // If RHSCst is [us]MAXINT, it is always false. Not handling
4159 // this can cause overflow.
4160 if (RHSCst->isMaxValue(false))
4161 return ReplaceInstUsesWith(I, LHS);
4162 return InsertRangeTest(LHSVal, LHSCst, AddOne(RHSCst), false,
4164 case ICmpInst::ICMP_SGT: // (X u< 13 | X s> 15) -> no change
4166 case ICmpInst::ICMP_NE: // (X u< 13 | X != 15) -> X != 15
4167 case ICmpInst::ICMP_ULT: // (X u< 13 | X u< 15) -> X u< 15
4168 return ReplaceInstUsesWith(I, RHS);
4169 case ICmpInst::ICMP_SLT: // (X u< 13 | X s< 15) -> no change
4173 case ICmpInst::ICMP_SLT:
4175 default: assert(0 && "Unknown integer condition code!");
4176 case ICmpInst::ICMP_EQ: // (X s< 13 | X == 14) -> no change
4178 case ICmpInst::ICMP_SGT: // (X s< 13 | X s> 15) ->(X-13) s> 2
4179 // If RHSCst is [us]MAXINT, it is always false. Not handling
4180 // this can cause overflow.
4181 if (RHSCst->isMaxValue(true))
4182 return ReplaceInstUsesWith(I, LHS);
4183 return InsertRangeTest(LHSVal, LHSCst, AddOne(RHSCst), true,
4185 case ICmpInst::ICMP_UGT: // (X s< 13 | X u> 15) -> no change
4187 case ICmpInst::ICMP_NE: // (X s< 13 | X != 15) -> X != 15
4188 case ICmpInst::ICMP_SLT: // (X s< 13 | X s< 15) -> X s< 15
4189 return ReplaceInstUsesWith(I, RHS);
4190 case ICmpInst::ICMP_ULT: // (X s< 13 | X u< 15) -> no change
4194 case ICmpInst::ICMP_UGT:
4196 default: assert(0 && "Unknown integer condition code!");
4197 case ICmpInst::ICMP_EQ: // (X u> 13 | X == 15) -> X u> 13
4198 case ICmpInst::ICMP_UGT: // (X u> 13 | X u> 15) -> X u> 13
4199 return ReplaceInstUsesWith(I, LHS);
4200 case ICmpInst::ICMP_SGT: // (X u> 13 | X s> 15) -> no change
4202 case ICmpInst::ICMP_NE: // (X u> 13 | X != 15) -> true
4203 case ICmpInst::ICMP_ULT: // (X u> 13 | X u< 15) -> true
4204 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4205 case ICmpInst::ICMP_SLT: // (X u> 13 | X s< 15) -> no change
4209 case ICmpInst::ICMP_SGT:
4211 default: assert(0 && "Unknown integer condition code!");
4212 case ICmpInst::ICMP_EQ: // (X s> 13 | X == 15) -> X > 13
4213 case ICmpInst::ICMP_SGT: // (X s> 13 | X s> 15) -> X > 13
4214 return ReplaceInstUsesWith(I, LHS);
4215 case ICmpInst::ICMP_UGT: // (X s> 13 | X u> 15) -> no change
4217 case ICmpInst::ICMP_NE: // (X s> 13 | X != 15) -> true
4218 case ICmpInst::ICMP_SLT: // (X s> 13 | X s< 15) -> true
4219 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4220 case ICmpInst::ICMP_ULT: // (X s> 13 | X u< 15) -> no change
4228 // fold (or (cast A), (cast B)) -> (cast (or A, B))
4229 if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) {
4230 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
4231 if (Op0C->getOpcode() == Op1C->getOpcode()) {// same cast kind ?
4232 const Type *SrcTy = Op0C->getOperand(0)->getType();
4233 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isInteger() &&
4234 // Only do this if the casts both really cause code to be generated.
4235 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
4237 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
4239 Instruction *NewOp = BinaryOperator::createOr(Op0C->getOperand(0),
4240 Op1C->getOperand(0),
4242 InsertNewInstBefore(NewOp, I);
4243 return CastInst::create(Op0C->getOpcode(), NewOp, I.getType());
4249 // (fcmp uno x, c) | (fcmp uno y, c) -> (fcmp uno x, y)
4250 if (FCmpInst *LHS = dyn_cast<FCmpInst>(I.getOperand(0))) {
4251 if (FCmpInst *RHS = dyn_cast<FCmpInst>(I.getOperand(1))) {
4252 if (LHS->getPredicate() == FCmpInst::FCMP_UNO &&
4253 RHS->getPredicate() == FCmpInst::FCMP_UNO)
4254 if (ConstantFP *LHSC = dyn_cast<ConstantFP>(LHS->getOperand(1)))
4255 if (ConstantFP *RHSC = dyn_cast<ConstantFP>(RHS->getOperand(1))) {
4256 // If either of the constants are nans, then the whole thing returns
4258 if (LHSC->getValueAPF().isNaN() || RHSC->getValueAPF().isNaN())
4259 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4261 // Otherwise, no need to compare the two constants, compare the
4263 return new FCmpInst(FCmpInst::FCMP_UNO, LHS->getOperand(0),
4264 RHS->getOperand(0));
4269 return Changed ? &I : 0;
4272 // XorSelf - Implements: X ^ X --> 0
4275 XorSelf(Value *rhs) : RHS(rhs) {}
4276 bool shouldApply(Value *LHS) const { return LHS == RHS; }
4277 Instruction *apply(BinaryOperator &Xor) const {
4283 Instruction *InstCombiner::visitXor(BinaryOperator &I) {
4284 bool Changed = SimplifyCommutative(I);
4285 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
4287 if (isa<UndefValue>(Op1))
4288 return ReplaceInstUsesWith(I, Op1); // X ^ undef -> undef
4290 // xor X, X = 0, even if X is nested in a sequence of Xor's.
4291 if (Instruction *Result = AssociativeOpt(I, XorSelf(Op1))) {
4292 assert(Result == &I && "AssociativeOpt didn't work?"); Result=Result;
4293 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
4296 // See if we can simplify any instructions used by the instruction whose sole
4297 // purpose is to compute bits we don't care about.
4298 if (!isa<VectorType>(I.getType())) {
4299 uint32_t BitWidth = cast<IntegerType>(I.getType())->getBitWidth();
4300 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
4301 if (SimplifyDemandedBits(&I, APInt::getAllOnesValue(BitWidth),
4302 KnownZero, KnownOne))
4304 } else if (isa<ConstantAggregateZero>(Op1)) {
4305 return ReplaceInstUsesWith(I, Op0); // X ^ <0,0> -> X
4308 // Is this a ~ operation?
4309 if (Value *NotOp = dyn_castNotVal(&I)) {
4310 // ~(~X & Y) --> (X | ~Y) - De Morgan's Law
4311 // ~(~X | Y) === (X & ~Y) - De Morgan's Law
4312 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(NotOp)) {
4313 if (Op0I->getOpcode() == Instruction::And ||
4314 Op0I->getOpcode() == Instruction::Or) {
4315 if (dyn_castNotVal(Op0I->getOperand(1))) Op0I->swapOperands();
4316 if (Value *Op0NotVal = dyn_castNotVal(Op0I->getOperand(0))) {
4318 BinaryOperator::createNot(Op0I->getOperand(1),
4319 Op0I->getOperand(1)->getName()+".not");
4320 InsertNewInstBefore(NotY, I);
4321 if (Op0I->getOpcode() == Instruction::And)
4322 return BinaryOperator::createOr(Op0NotVal, NotY);
4324 return BinaryOperator::createAnd(Op0NotVal, NotY);
4331 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
4332 // xor (cmp A, B), true = not (cmp A, B) = !cmp A, B
4333 if (RHS == ConstantInt::getTrue() && Op0->hasOneUse()) {
4334 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Op0))
4335 return new ICmpInst(ICI->getInversePredicate(),
4336 ICI->getOperand(0), ICI->getOperand(1));
4338 if (FCmpInst *FCI = dyn_cast<FCmpInst>(Op0))
4339 return new FCmpInst(FCI->getInversePredicate(),
4340 FCI->getOperand(0), FCI->getOperand(1));
4343 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0)) {
4344 // ~(c-X) == X-c-1 == X+(-c-1)
4345 if (Op0I->getOpcode() == Instruction::Sub && RHS->isAllOnesValue())
4346 if (Constant *Op0I0C = dyn_cast<Constant>(Op0I->getOperand(0))) {
4347 Constant *NegOp0I0C = ConstantExpr::getNeg(Op0I0C);
4348 Constant *ConstantRHS = ConstantExpr::getSub(NegOp0I0C,
4349 ConstantInt::get(I.getType(), 1));
4350 return BinaryOperator::createAdd(Op0I->getOperand(1), ConstantRHS);
4353 if (ConstantInt *Op0CI = dyn_cast<ConstantInt>(Op0I->getOperand(1)))
4354 if (Op0I->getOpcode() == Instruction::Add) {
4355 // ~(X-c) --> (-c-1)-X
4356 if (RHS->isAllOnesValue()) {
4357 Constant *NegOp0CI = ConstantExpr::getNeg(Op0CI);
4358 return BinaryOperator::createSub(
4359 ConstantExpr::getSub(NegOp0CI,
4360 ConstantInt::get(I.getType(), 1)),
4361 Op0I->getOperand(0));
4362 } else if (RHS->getValue().isSignBit()) {
4363 // (X + C) ^ signbit -> (X + C + signbit)
4364 Constant *C = ConstantInt::get(RHS->getValue() + Op0CI->getValue());
4365 return BinaryOperator::createAdd(Op0I->getOperand(0), C);
4368 } else if (Op0I->getOpcode() == Instruction::Or) {
4369 // (X|C1)^C2 -> X^(C1|C2) iff X&~C1 == 0
4370 if (MaskedValueIsZero(Op0I->getOperand(0), Op0CI->getValue())) {
4371 Constant *NewRHS = ConstantExpr::getOr(Op0CI, RHS);
4372 // Anything in both C1 and C2 is known to be zero, remove it from
4374 Constant *CommonBits = And(Op0CI, RHS);
4375 NewRHS = ConstantExpr::getAnd(NewRHS,
4376 ConstantExpr::getNot(CommonBits));
4377 AddToWorkList(Op0I);
4378 I.setOperand(0, Op0I->getOperand(0));
4379 I.setOperand(1, NewRHS);
4385 // Try to fold constant and into select arguments.
4386 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
4387 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
4389 if (isa<PHINode>(Op0))
4390 if (Instruction *NV = FoldOpIntoPhi(I))
4394 if (Value *X = dyn_castNotVal(Op0)) // ~A ^ A == -1
4396 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
4398 if (Value *X = dyn_castNotVal(Op1)) // A ^ ~A == -1
4400 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
4403 BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1);
4406 if (match(Op1I, m_Or(m_Value(A), m_Value(B)))) {
4407 if (A == Op0) { // B^(B|A) == (A|B)^B
4408 Op1I->swapOperands();
4410 std::swap(Op0, Op1);
4411 } else if (B == Op0) { // B^(A|B) == (A|B)^B
4412 I.swapOperands(); // Simplified below.
4413 std::swap(Op0, Op1);
4415 } else if (match(Op1I, m_Xor(m_Value(A), m_Value(B)))) {
4416 if (Op0 == A) // A^(A^B) == B
4417 return ReplaceInstUsesWith(I, B);
4418 else if (Op0 == B) // A^(B^A) == B
4419 return ReplaceInstUsesWith(I, A);
4420 } else if (match(Op1I, m_And(m_Value(A), m_Value(B))) && Op1I->hasOneUse()){
4421 if (A == Op0) { // A^(A&B) -> A^(B&A)
4422 Op1I->swapOperands();
4425 if (B == Op0) { // A^(B&A) -> (B&A)^A
4426 I.swapOperands(); // Simplified below.
4427 std::swap(Op0, Op1);
4432 BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0);
4435 if (match(Op0I, m_Or(m_Value(A), m_Value(B))) && Op0I->hasOneUse()) {
4436 if (A == Op1) // (B|A)^B == (A|B)^B
4438 if (B == Op1) { // (A|B)^B == A & ~B
4440 InsertNewInstBefore(BinaryOperator::createNot(Op1, "tmp"), I);
4441 return BinaryOperator::createAnd(A, NotB);
4443 } else if (match(Op0I, m_Xor(m_Value(A), m_Value(B)))) {
4444 if (Op1 == A) // (A^B)^A == B
4445 return ReplaceInstUsesWith(I, B);
4446 else if (Op1 == B) // (B^A)^A == B
4447 return ReplaceInstUsesWith(I, A);
4448 } else if (match(Op0I, m_And(m_Value(A), m_Value(B))) && Op0I->hasOneUse()){
4449 if (A == Op1) // (A&B)^A -> (B&A)^A
4451 if (B == Op1 && // (B&A)^A == ~B & A
4452 !isa<ConstantInt>(Op1)) { // Canonical form is (B&C)^C
4454 InsertNewInstBefore(BinaryOperator::createNot(A, "tmp"), I);
4455 return BinaryOperator::createAnd(N, Op1);
4460 // (X >> Z) ^ (Y >> Z) -> (X^Y) >> Z for all shifts.
4461 if (Op0I && Op1I && Op0I->isShift() &&
4462 Op0I->getOpcode() == Op1I->getOpcode() &&
4463 Op0I->getOperand(1) == Op1I->getOperand(1) &&
4464 (Op1I->hasOneUse() || Op1I->hasOneUse())) {
4465 Instruction *NewOp =
4466 InsertNewInstBefore(BinaryOperator::createXor(Op0I->getOperand(0),
4467 Op1I->getOperand(0),
4468 Op0I->getName()), I);
4469 return BinaryOperator::create(Op1I->getOpcode(), NewOp,
4470 Op1I->getOperand(1));
4474 Value *A, *B, *C, *D;
4475 // (A & B)^(A | B) -> A ^ B
4476 if (match(Op0I, m_And(m_Value(A), m_Value(B))) &&
4477 match(Op1I, m_Or(m_Value(C), m_Value(D)))) {
4478 if ((A == C && B == D) || (A == D && B == C))
4479 return BinaryOperator::createXor(A, B);
4481 // (A | B)^(A & B) -> A ^ B
4482 if (match(Op0I, m_Or(m_Value(A), m_Value(B))) &&
4483 match(Op1I, m_And(m_Value(C), m_Value(D)))) {
4484 if ((A == C && B == D) || (A == D && B == C))
4485 return BinaryOperator::createXor(A, B);
4489 if ((Op0I->hasOneUse() || Op1I->hasOneUse()) &&
4490 match(Op0I, m_And(m_Value(A), m_Value(B))) &&
4491 match(Op1I, m_And(m_Value(C), m_Value(D)))) {
4492 // (X & Y)^(X & Y) -> (Y^Z) & X
4493 Value *X = 0, *Y = 0, *Z = 0;
4495 X = A, Y = B, Z = D;
4497 X = A, Y = B, Z = C;
4499 X = B, Y = A, Z = D;
4501 X = B, Y = A, Z = C;
4504 Instruction *NewOp =
4505 InsertNewInstBefore(BinaryOperator::createXor(Y, Z, Op0->getName()), I);
4506 return BinaryOperator::createAnd(NewOp, X);
4511 // (icmp1 A, B) ^ (icmp2 A, B) --> (icmp3 A, B)
4512 if (ICmpInst *RHS = dyn_cast<ICmpInst>(I.getOperand(1)))
4513 if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS)))
4516 // fold (xor (cast A), (cast B)) -> (cast (xor A, B))
4517 if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) {
4518 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
4519 if (Op0C->getOpcode() == Op1C->getOpcode()) { // same cast kind?
4520 const Type *SrcTy = Op0C->getOperand(0)->getType();
4521 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isInteger() &&
4522 // Only do this if the casts both really cause code to be generated.
4523 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
4525 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
4527 Instruction *NewOp = BinaryOperator::createXor(Op0C->getOperand(0),
4528 Op1C->getOperand(0),
4530 InsertNewInstBefore(NewOp, I);
4531 return CastInst::create(Op0C->getOpcode(), NewOp, I.getType());
4535 return Changed ? &I : 0;
4538 /// AddWithOverflow - Compute Result = In1+In2, returning true if the result
4539 /// overflowed for this type.
4540 static bool AddWithOverflow(ConstantInt *&Result, ConstantInt *In1,
4541 ConstantInt *In2, bool IsSigned = false) {
4542 Result = cast<ConstantInt>(Add(In1, In2));
4545 if (In2->getValue().isNegative())
4546 return Result->getValue().sgt(In1->getValue());
4548 return Result->getValue().slt(In1->getValue());
4550 return Result->getValue().ult(In1->getValue());
4553 /// EmitGEPOffset - Given a getelementptr instruction/constantexpr, emit the
4554 /// code necessary to compute the offset from the base pointer (without adding
4555 /// in the base pointer). Return the result as a signed integer of intptr size.
4556 static Value *EmitGEPOffset(User *GEP, Instruction &I, InstCombiner &IC) {
4557 TargetData &TD = IC.getTargetData();
4558 gep_type_iterator GTI = gep_type_begin(GEP);
4559 const Type *IntPtrTy = TD.getIntPtrType();
4560 Value *Result = Constant::getNullValue(IntPtrTy);
4562 // Build a mask for high order bits.
4563 unsigned IntPtrWidth = TD.getPointerSize()*8;
4564 uint64_t PtrSizeMask = ~0ULL >> (64-IntPtrWidth);
4566 for (unsigned i = 1, e = GEP->getNumOperands(); i != e; ++i, ++GTI) {
4567 Value *Op = GEP->getOperand(i);
4568 uint64_t Size = TD.getABITypeSize(GTI.getIndexedType()) & PtrSizeMask;
4569 if (ConstantInt *OpC = dyn_cast<ConstantInt>(Op)) {
4570 if (OpC->isZero()) continue;
4572 // Handle a struct index, which adds its field offset to the pointer.
4573 if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
4574 Size = TD.getStructLayout(STy)->getElementOffset(OpC->getZExtValue());
4576 if (ConstantInt *RC = dyn_cast<ConstantInt>(Result))
4577 Result = ConstantInt::get(RC->getValue() + APInt(IntPtrWidth, Size));
4579 Result = IC.InsertNewInstBefore(
4580 BinaryOperator::createAdd(Result,
4581 ConstantInt::get(IntPtrTy, Size),
4582 GEP->getName()+".offs"), I);
4586 Constant *Scale = ConstantInt::get(IntPtrTy, Size);
4587 Constant *OC = ConstantExpr::getIntegerCast(OpC, IntPtrTy, true /*SExt*/);
4588 Scale = ConstantExpr::getMul(OC, Scale);
4589 if (Constant *RC = dyn_cast<Constant>(Result))
4590 Result = ConstantExpr::getAdd(RC, Scale);
4592 // Emit an add instruction.
4593 Result = IC.InsertNewInstBefore(
4594 BinaryOperator::createAdd(Result, Scale,
4595 GEP->getName()+".offs"), I);
4599 // Convert to correct type.
4600 if (Op->getType() != IntPtrTy) {
4601 if (Constant *OpC = dyn_cast<Constant>(Op))
4602 Op = ConstantExpr::getSExt(OpC, IntPtrTy);
4604 Op = IC.InsertNewInstBefore(new SExtInst(Op, IntPtrTy,
4605 Op->getName()+".c"), I);
4608 Constant *Scale = ConstantInt::get(IntPtrTy, Size);
4609 if (Constant *OpC = dyn_cast<Constant>(Op))
4610 Op = ConstantExpr::getMul(OpC, Scale);
4611 else // We'll let instcombine(mul) convert this to a shl if possible.
4612 Op = IC.InsertNewInstBefore(BinaryOperator::createMul(Op, Scale,
4613 GEP->getName()+".idx"), I);
4616 // Emit an add instruction.
4617 if (isa<Constant>(Op) && isa<Constant>(Result))
4618 Result = ConstantExpr::getAdd(cast<Constant>(Op),
4619 cast<Constant>(Result));
4621 Result = IC.InsertNewInstBefore(BinaryOperator::createAdd(Op, Result,
4622 GEP->getName()+".offs"), I);
4627 /// FoldGEPICmp - Fold comparisons between a GEP instruction and something
4628 /// else. At this point we know that the GEP is on the LHS of the comparison.
4629 Instruction *InstCombiner::FoldGEPICmp(User *GEPLHS, Value *RHS,
4630 ICmpInst::Predicate Cond,
4632 assert(dyn_castGetElementPtr(GEPLHS) && "LHS is not a getelementptr!");
4634 if (CastInst *CI = dyn_cast<CastInst>(RHS))
4635 if (isa<PointerType>(CI->getOperand(0)->getType()))
4636 RHS = CI->getOperand(0);
4638 Value *PtrBase = GEPLHS->getOperand(0);
4639 if (PtrBase == RHS) {
4640 // ((gep Ptr, OFFSET) cmp Ptr) ---> (OFFSET cmp 0).
4641 // This transformation is valid because we know pointers can't overflow.
4642 Value *Offset = EmitGEPOffset(GEPLHS, I, *this);
4643 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), Offset,
4644 Constant::getNullValue(Offset->getType()));
4645 } else if (User *GEPRHS = dyn_castGetElementPtr(RHS)) {
4646 // If the base pointers are different, but the indices are the same, just
4647 // compare the base pointer.
4648 if (PtrBase != GEPRHS->getOperand(0)) {
4649 bool IndicesTheSame = GEPLHS->getNumOperands()==GEPRHS->getNumOperands();
4650 IndicesTheSame &= GEPLHS->getOperand(0)->getType() ==
4651 GEPRHS->getOperand(0)->getType();
4653 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
4654 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
4655 IndicesTheSame = false;
4659 // If all indices are the same, just compare the base pointers.
4661 return new ICmpInst(ICmpInst::getSignedPredicate(Cond),
4662 GEPLHS->getOperand(0), GEPRHS->getOperand(0));
4664 // Otherwise, the base pointers are different and the indices are
4665 // different, bail out.
4669 // If one of the GEPs has all zero indices, recurse.
4670 bool AllZeros = true;
4671 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
4672 if (!isa<Constant>(GEPLHS->getOperand(i)) ||
4673 !cast<Constant>(GEPLHS->getOperand(i))->isNullValue()) {
4678 return FoldGEPICmp(GEPRHS, GEPLHS->getOperand(0),
4679 ICmpInst::getSwappedPredicate(Cond), I);
4681 // If the other GEP has all zero indices, recurse.
4683 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
4684 if (!isa<Constant>(GEPRHS->getOperand(i)) ||
4685 !cast<Constant>(GEPRHS->getOperand(i))->isNullValue()) {
4690 return FoldGEPICmp(GEPLHS, GEPRHS->getOperand(0), Cond, I);
4692 if (GEPLHS->getNumOperands() == GEPRHS->getNumOperands()) {
4693 // If the GEPs only differ by one index, compare it.
4694 unsigned NumDifferences = 0; // Keep track of # differences.
4695 unsigned DiffOperand = 0; // The operand that differs.
4696 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
4697 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
4698 if (GEPLHS->getOperand(i)->getType()->getPrimitiveSizeInBits() !=
4699 GEPRHS->getOperand(i)->getType()->getPrimitiveSizeInBits()) {
4700 // Irreconcilable differences.
4704 if (NumDifferences++) break;
4709 if (NumDifferences == 0) // SAME GEP?
4710 return ReplaceInstUsesWith(I, // No comparison is needed here.
4711 ConstantInt::get(Type::Int1Ty,
4712 isTrueWhenEqual(Cond)));
4714 else if (NumDifferences == 1) {
4715 Value *LHSV = GEPLHS->getOperand(DiffOperand);
4716 Value *RHSV = GEPRHS->getOperand(DiffOperand);
4717 // Make sure we do a signed comparison here.
4718 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), LHSV, RHSV);
4722 // Only lower this if the icmp is the only user of the GEP or if we expect
4723 // the result to fold to a constant!
4724 if ((isa<ConstantExpr>(GEPLHS) || GEPLHS->hasOneUse()) &&
4725 (isa<ConstantExpr>(GEPRHS) || GEPRHS->hasOneUse())) {
4726 // ((gep Ptr, OFFSET1) cmp (gep Ptr, OFFSET2) ---> (OFFSET1 cmp OFFSET2)
4727 Value *L = EmitGEPOffset(GEPLHS, I, *this);
4728 Value *R = EmitGEPOffset(GEPRHS, I, *this);
4729 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), L, R);
4735 Instruction *InstCombiner::visitFCmpInst(FCmpInst &I) {
4736 bool Changed = SimplifyCompare(I);
4737 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
4739 // Fold trivial predicates.
4740 if (I.getPredicate() == FCmpInst::FCMP_FALSE)
4741 return ReplaceInstUsesWith(I, Constant::getNullValue(Type::Int1Ty));
4742 if (I.getPredicate() == FCmpInst::FCMP_TRUE)
4743 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 1));
4745 // Simplify 'fcmp pred X, X'
4747 switch (I.getPredicate()) {
4748 default: assert(0 && "Unknown predicate!");
4749 case FCmpInst::FCMP_UEQ: // True if unordered or equal
4750 case FCmpInst::FCMP_UGE: // True if unordered, greater than, or equal
4751 case FCmpInst::FCMP_ULE: // True if unordered, less than, or equal
4752 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 1));
4753 case FCmpInst::FCMP_OGT: // True if ordered and greater than
4754 case FCmpInst::FCMP_OLT: // True if ordered and less than
4755 case FCmpInst::FCMP_ONE: // True if ordered and operands are unequal
4756 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 0));
4758 case FCmpInst::FCMP_UNO: // True if unordered: isnan(X) | isnan(Y)
4759 case FCmpInst::FCMP_ULT: // True if unordered or less than
4760 case FCmpInst::FCMP_UGT: // True if unordered or greater than
4761 case FCmpInst::FCMP_UNE: // True if unordered or not equal
4762 // Canonicalize these to be 'fcmp uno %X, 0.0'.
4763 I.setPredicate(FCmpInst::FCMP_UNO);
4764 I.setOperand(1, Constant::getNullValue(Op0->getType()));
4767 case FCmpInst::FCMP_ORD: // True if ordered (no nans)
4768 case FCmpInst::FCMP_OEQ: // True if ordered and equal
4769 case FCmpInst::FCMP_OGE: // True if ordered and greater than or equal
4770 case FCmpInst::FCMP_OLE: // True if ordered and less than or equal
4771 // Canonicalize these to be 'fcmp ord %X, 0.0'.
4772 I.setPredicate(FCmpInst::FCMP_ORD);
4773 I.setOperand(1, Constant::getNullValue(Op0->getType()));
4778 if (isa<UndefValue>(Op1)) // fcmp pred X, undef -> undef
4779 return ReplaceInstUsesWith(I, UndefValue::get(Type::Int1Ty));
4781 // Handle fcmp with constant RHS
4782 if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
4783 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
4784 switch (LHSI->getOpcode()) {
4785 case Instruction::PHI:
4786 if (Instruction *NV = FoldOpIntoPhi(I))
4789 case Instruction::Select:
4790 // If either operand of the select is a constant, we can fold the
4791 // comparison into the select arms, which will cause one to be
4792 // constant folded and the select turned into a bitwise or.
4793 Value *Op1 = 0, *Op2 = 0;
4794 if (LHSI->hasOneUse()) {
4795 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1))) {
4796 // Fold the known value into the constant operand.
4797 Op1 = ConstantExpr::getCompare(I.getPredicate(), C, RHSC);
4798 // Insert a new FCmp of the other select operand.
4799 Op2 = InsertNewInstBefore(new FCmpInst(I.getPredicate(),
4800 LHSI->getOperand(2), RHSC,
4802 } else if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2))) {
4803 // Fold the known value into the constant operand.
4804 Op2 = ConstantExpr::getCompare(I.getPredicate(), C, RHSC);
4805 // Insert a new FCmp of the other select operand.
4806 Op1 = InsertNewInstBefore(new FCmpInst(I.getPredicate(),
4807 LHSI->getOperand(1), RHSC,
4813 return new SelectInst(LHSI->getOperand(0), Op1, Op2);
4818 return Changed ? &I : 0;
4821 Instruction *InstCombiner::visitICmpInst(ICmpInst &I) {
4822 bool Changed = SimplifyCompare(I);
4823 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
4824 const Type *Ty = Op0->getType();
4828 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
4829 isTrueWhenEqual(I)));
4831 if (isa<UndefValue>(Op1)) // X icmp undef -> undef
4832 return ReplaceInstUsesWith(I, UndefValue::get(Type::Int1Ty));
4834 // icmp <global/alloca*/null>, <global/alloca*/null> - Global/Stack value
4835 // addresses never equal each other! We already know that Op0 != Op1.
4836 if ((isa<GlobalValue>(Op0) || isa<AllocaInst>(Op0) ||
4837 isa<ConstantPointerNull>(Op0)) &&
4838 (isa<GlobalValue>(Op1) || isa<AllocaInst>(Op1) ||
4839 isa<ConstantPointerNull>(Op1)))
4840 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
4841 !isTrueWhenEqual(I)));
4843 // icmp's with boolean values can always be turned into bitwise operations
4844 if (Ty == Type::Int1Ty) {
4845 switch (I.getPredicate()) {
4846 default: assert(0 && "Invalid icmp instruction!");
4847 case ICmpInst::ICMP_EQ: { // icmp eq bool %A, %B -> ~(A^B)
4848 Instruction *Xor = BinaryOperator::createXor(Op0, Op1, I.getName()+"tmp");
4849 InsertNewInstBefore(Xor, I);
4850 return BinaryOperator::createNot(Xor);
4852 case ICmpInst::ICMP_NE: // icmp eq bool %A, %B -> A^B
4853 return BinaryOperator::createXor(Op0, Op1);
4855 case ICmpInst::ICMP_UGT:
4856 case ICmpInst::ICMP_SGT:
4857 std::swap(Op0, Op1); // Change icmp gt -> icmp lt
4859 case ICmpInst::ICMP_ULT:
4860 case ICmpInst::ICMP_SLT: { // icmp lt bool A, B -> ~X & Y
4861 Instruction *Not = BinaryOperator::createNot(Op0, I.getName()+"tmp");
4862 InsertNewInstBefore(Not, I);
4863 return BinaryOperator::createAnd(Not, Op1);
4865 case ICmpInst::ICMP_UGE:
4866 case ICmpInst::ICMP_SGE:
4867 std::swap(Op0, Op1); // Change icmp ge -> icmp le
4869 case ICmpInst::ICMP_ULE:
4870 case ICmpInst::ICMP_SLE: { // icmp le bool %A, %B -> ~A | B
4871 Instruction *Not = BinaryOperator::createNot(Op0, I.getName()+"tmp");
4872 InsertNewInstBefore(Not, I);
4873 return BinaryOperator::createOr(Not, Op1);
4878 // See if we are doing a comparison between a constant and an instruction that
4879 // can be folded into the comparison.
4880 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
4883 // (icmp ne/eq (sub A B) 0) -> (icmp ne/eq A, B)
4884 if (I.isEquality() && CI->isNullValue() &&
4885 match(Op0, m_Sub(m_Value(A), m_Value(B)))) {
4886 // (icmp cond A B) if cond is equality
4887 return new ICmpInst(I.getPredicate(), A, B);
4890 switch (I.getPredicate()) {
4892 case ICmpInst::ICMP_ULT: // A <u MIN -> FALSE
4893 if (CI->isMinValue(false))
4894 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4895 if (CI->isMaxValue(false)) // A <u MAX -> A != MAX
4896 return new ICmpInst(ICmpInst::ICMP_NE, Op0,Op1);
4897 if (isMinValuePlusOne(CI,false)) // A <u MIN+1 -> A == MIN
4898 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, SubOne(CI));
4899 // (x <u 2147483648) -> (x >s -1) -> true if sign bit clear
4900 if (CI->isMinValue(true))
4901 return new ICmpInst(ICmpInst::ICMP_SGT, Op0,
4902 ConstantInt::getAllOnesValue(Op0->getType()));
4906 case ICmpInst::ICMP_SLT:
4907 if (CI->isMinValue(true)) // A <s MIN -> FALSE
4908 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4909 if (CI->isMaxValue(true)) // A <s MAX -> A != MAX
4910 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
4911 if (isMinValuePlusOne(CI,true)) // A <s MIN+1 -> A == MIN
4912 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, SubOne(CI));
4915 case ICmpInst::ICMP_UGT:
4916 if (CI->isMaxValue(false)) // A >u MAX -> FALSE
4917 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4918 if (CI->isMinValue(false)) // A >u MIN -> A != MIN
4919 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
4920 if (isMaxValueMinusOne(CI, false)) // A >u MAX-1 -> A == MAX
4921 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, AddOne(CI));
4923 // (x >u 2147483647) -> (x <s 0) -> true if sign bit set
4924 if (CI->isMaxValue(true))
4925 return new ICmpInst(ICmpInst::ICMP_SLT, Op0,
4926 ConstantInt::getNullValue(Op0->getType()));
4929 case ICmpInst::ICMP_SGT:
4930 if (CI->isMaxValue(true)) // A >s MAX -> FALSE
4931 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4932 if (CI->isMinValue(true)) // A >s MIN -> A != MIN
4933 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
4934 if (isMaxValueMinusOne(CI, true)) // A >s MAX-1 -> A == MAX
4935 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, AddOne(CI));
4938 case ICmpInst::ICMP_ULE:
4939 if (CI->isMaxValue(false)) // A <=u MAX -> TRUE
4940 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4941 if (CI->isMinValue(false)) // A <=u MIN -> A == MIN
4942 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
4943 if (isMaxValueMinusOne(CI,false)) // A <=u MAX-1 -> A != MAX
4944 return new ICmpInst(ICmpInst::ICMP_NE, Op0, AddOne(CI));
4947 case ICmpInst::ICMP_SLE:
4948 if (CI->isMaxValue(true)) // A <=s MAX -> TRUE
4949 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4950 if (CI->isMinValue(true)) // A <=s MIN -> A == MIN
4951 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
4952 if (isMaxValueMinusOne(CI,true)) // A <=s MAX-1 -> A != MAX
4953 return new ICmpInst(ICmpInst::ICMP_NE, Op0, AddOne(CI));
4956 case ICmpInst::ICMP_UGE:
4957 if (CI->isMinValue(false)) // A >=u MIN -> TRUE
4958 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4959 if (CI->isMaxValue(false)) // A >=u MAX -> A == MAX
4960 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
4961 if (isMinValuePlusOne(CI,false)) // A >=u MIN-1 -> A != MIN
4962 return new ICmpInst(ICmpInst::ICMP_NE, Op0, SubOne(CI));
4965 case ICmpInst::ICMP_SGE:
4966 if (CI->isMinValue(true)) // A >=s MIN -> TRUE
4967 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4968 if (CI->isMaxValue(true)) // A >=s MAX -> A == MAX
4969 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
4970 if (isMinValuePlusOne(CI,true)) // A >=s MIN-1 -> A != MIN
4971 return new ICmpInst(ICmpInst::ICMP_NE, Op0, SubOne(CI));
4975 // If we still have a icmp le or icmp ge instruction, turn it into the
4976 // appropriate icmp lt or icmp gt instruction. Since the border cases have
4977 // already been handled above, this requires little checking.
4979 switch (I.getPredicate()) {
4981 case ICmpInst::ICMP_ULE:
4982 return new ICmpInst(ICmpInst::ICMP_ULT, Op0, AddOne(CI));
4983 case ICmpInst::ICMP_SLE:
4984 return new ICmpInst(ICmpInst::ICMP_SLT, Op0, AddOne(CI));
4985 case ICmpInst::ICMP_UGE:
4986 return new ICmpInst( ICmpInst::ICMP_UGT, Op0, SubOne(CI));
4987 case ICmpInst::ICMP_SGE:
4988 return new ICmpInst(ICmpInst::ICMP_SGT, Op0, SubOne(CI));
4991 // See if we can fold the comparison based on bits known to be zero or one
4992 // in the input. If this comparison is a normal comparison, it demands all
4993 // bits, if it is a sign bit comparison, it only demands the sign bit.
4996 bool isSignBit = isSignBitCheck(I.getPredicate(), CI, UnusedBit);
4998 uint32_t BitWidth = cast<IntegerType>(Ty)->getBitWidth();
4999 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
5000 if (SimplifyDemandedBits(Op0,
5001 isSignBit ? APInt::getSignBit(BitWidth)
5002 : APInt::getAllOnesValue(BitWidth),
5003 KnownZero, KnownOne, 0))
5006 // Given the known and unknown bits, compute a range that the LHS could be
5008 if ((KnownOne | KnownZero) != 0) {
5009 // Compute the Min, Max and RHS values based on the known bits. For the
5010 // EQ and NE we use unsigned values.
5011 APInt Min(BitWidth, 0), Max(BitWidth, 0);
5012 const APInt& RHSVal = CI->getValue();
5013 if (ICmpInst::isSignedPredicate(I.getPredicate())) {
5014 ComputeSignedMinMaxValuesFromKnownBits(Ty, KnownZero, KnownOne, Min,
5017 ComputeUnsignedMinMaxValuesFromKnownBits(Ty, KnownZero, KnownOne, Min,
5020 switch (I.getPredicate()) { // LE/GE have been folded already.
5021 default: assert(0 && "Unknown icmp opcode!");
5022 case ICmpInst::ICMP_EQ:
5023 if (Max.ult(RHSVal) || Min.ugt(RHSVal))
5024 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5026 case ICmpInst::ICMP_NE:
5027 if (Max.ult(RHSVal) || Min.ugt(RHSVal))
5028 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5030 case ICmpInst::ICMP_ULT:
5031 if (Max.ult(RHSVal))
5032 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5033 if (Min.uge(RHSVal))
5034 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5036 case ICmpInst::ICMP_UGT:
5037 if (Min.ugt(RHSVal))
5038 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5039 if (Max.ule(RHSVal))
5040 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5042 case ICmpInst::ICMP_SLT:
5043 if (Max.slt(RHSVal))
5044 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5045 if (Min.sgt(RHSVal))
5046 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5048 case ICmpInst::ICMP_SGT:
5049 if (Min.sgt(RHSVal))
5050 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5051 if (Max.sle(RHSVal))
5052 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5057 // Since the RHS is a ConstantInt (CI), if the left hand side is an
5058 // instruction, see if that instruction also has constants so that the
5059 // instruction can be folded into the icmp
5060 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
5061 if (Instruction *Res = visitICmpInstWithInstAndIntCst(I, LHSI, CI))
5065 // Handle icmp with constant (but not simple integer constant) RHS
5066 if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
5067 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
5068 switch (LHSI->getOpcode()) {
5069 case Instruction::GetElementPtr:
5070 if (RHSC->isNullValue()) {
5071 // icmp pred GEP (P, int 0, int 0, int 0), null -> icmp pred P, null
5072 bool isAllZeros = true;
5073 for (unsigned i = 1, e = LHSI->getNumOperands(); i != e; ++i)
5074 if (!isa<Constant>(LHSI->getOperand(i)) ||
5075 !cast<Constant>(LHSI->getOperand(i))->isNullValue()) {
5080 return new ICmpInst(I.getPredicate(), LHSI->getOperand(0),
5081 Constant::getNullValue(LHSI->getOperand(0)->getType()));
5085 case Instruction::PHI:
5086 if (Instruction *NV = FoldOpIntoPhi(I))
5089 case Instruction::Select: {
5090 // If either operand of the select is a constant, we can fold the
5091 // comparison into the select arms, which will cause one to be
5092 // constant folded and the select turned into a bitwise or.
5093 Value *Op1 = 0, *Op2 = 0;
5094 if (LHSI->hasOneUse()) {
5095 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1))) {
5096 // Fold the known value into the constant operand.
5097 Op1 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
5098 // Insert a new ICmp of the other select operand.
5099 Op2 = InsertNewInstBefore(new ICmpInst(I.getPredicate(),
5100 LHSI->getOperand(2), RHSC,
5102 } else if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2))) {
5103 // Fold the known value into the constant operand.
5104 Op2 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
5105 // Insert a new ICmp of the other select operand.
5106 Op1 = InsertNewInstBefore(new ICmpInst(I.getPredicate(),
5107 LHSI->getOperand(1), RHSC,
5113 return new SelectInst(LHSI->getOperand(0), Op1, Op2);
5116 case Instruction::Malloc:
5117 // If we have (malloc != null), and if the malloc has a single use, we
5118 // can assume it is successful and remove the malloc.
5119 if (LHSI->hasOneUse() && isa<ConstantPointerNull>(RHSC)) {
5120 AddToWorkList(LHSI);
5121 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
5122 !isTrueWhenEqual(I)));
5128 // If we can optimize a 'icmp GEP, P' or 'icmp P, GEP', do so now.
5129 if (User *GEP = dyn_castGetElementPtr(Op0))
5130 if (Instruction *NI = FoldGEPICmp(GEP, Op1, I.getPredicate(), I))
5132 if (User *GEP = dyn_castGetElementPtr(Op1))
5133 if (Instruction *NI = FoldGEPICmp(GEP, Op0,
5134 ICmpInst::getSwappedPredicate(I.getPredicate()), I))
5137 // Test to see if the operands of the icmp are casted versions of other
5138 // values. If the ptr->ptr cast can be stripped off both arguments, we do so
5140 if (BitCastInst *CI = dyn_cast<BitCastInst>(Op0)) {
5141 if (isa<PointerType>(Op0->getType()) &&
5142 (isa<Constant>(Op1) || isa<BitCastInst>(Op1))) {
5143 // We keep moving the cast from the left operand over to the right
5144 // operand, where it can often be eliminated completely.
5145 Op0 = CI->getOperand(0);
5147 // If operand #1 is a bitcast instruction, it must also be a ptr->ptr cast
5148 // so eliminate it as well.
5149 if (BitCastInst *CI2 = dyn_cast<BitCastInst>(Op1))
5150 Op1 = CI2->getOperand(0);
5152 // If Op1 is a constant, we can fold the cast into the constant.
5153 if (Op0->getType() != Op1->getType())
5154 if (Constant *Op1C = dyn_cast<Constant>(Op1)) {
5155 Op1 = ConstantExpr::getBitCast(Op1C, Op0->getType());
5157 // Otherwise, cast the RHS right before the icmp
5158 Op1 = InsertBitCastBefore(Op1, Op0->getType(), I);
5160 return new ICmpInst(I.getPredicate(), Op0, Op1);
5164 if (isa<CastInst>(Op0)) {
5165 // Handle the special case of: icmp (cast bool to X), <cst>
5166 // This comes up when you have code like
5169 // For generality, we handle any zero-extension of any operand comparison
5170 // with a constant or another cast from the same type.
5171 if (isa<ConstantInt>(Op1) || isa<CastInst>(Op1))
5172 if (Instruction *R = visitICmpInstWithCastAndCast(I))
5176 if (I.isEquality()) {
5177 Value *A, *B, *C, *D;
5178 if (match(Op0, m_Xor(m_Value(A), m_Value(B)))) {
5179 if (A == Op1 || B == Op1) { // (A^B) == A -> B == 0
5180 Value *OtherVal = A == Op1 ? B : A;
5181 return new ICmpInst(I.getPredicate(), OtherVal,
5182 Constant::getNullValue(A->getType()));
5185 if (match(Op1, m_Xor(m_Value(C), m_Value(D)))) {
5186 // A^c1 == C^c2 --> A == C^(c1^c2)
5187 if (ConstantInt *C1 = dyn_cast<ConstantInt>(B))
5188 if (ConstantInt *C2 = dyn_cast<ConstantInt>(D))
5189 if (Op1->hasOneUse()) {
5190 Constant *NC = ConstantInt::get(C1->getValue() ^ C2->getValue());
5191 Instruction *Xor = BinaryOperator::createXor(C, NC, "tmp");
5192 return new ICmpInst(I.getPredicate(), A,
5193 InsertNewInstBefore(Xor, I));
5196 // A^B == A^D -> B == D
5197 if (A == C) return new ICmpInst(I.getPredicate(), B, D);
5198 if (A == D) return new ICmpInst(I.getPredicate(), B, C);
5199 if (B == C) return new ICmpInst(I.getPredicate(), A, D);
5200 if (B == D) return new ICmpInst(I.getPredicate(), A, C);
5204 if (match(Op1, m_Xor(m_Value(A), m_Value(B))) &&
5205 (A == Op0 || B == Op0)) {
5206 // A == (A^B) -> B == 0
5207 Value *OtherVal = A == Op0 ? B : A;
5208 return new ICmpInst(I.getPredicate(), OtherVal,
5209 Constant::getNullValue(A->getType()));
5211 if (match(Op0, m_Sub(m_Value(A), m_Value(B))) && A == Op1) {
5212 // (A-B) == A -> B == 0
5213 return new ICmpInst(I.getPredicate(), B,
5214 Constant::getNullValue(B->getType()));
5216 if (match(Op1, m_Sub(m_Value(A), m_Value(B))) && A == Op0) {
5217 // A == (A-B) -> B == 0
5218 return new ICmpInst(I.getPredicate(), B,
5219 Constant::getNullValue(B->getType()));
5222 // (X&Z) == (Y&Z) -> (X^Y) & Z == 0
5223 if (Op0->hasOneUse() && Op1->hasOneUse() &&
5224 match(Op0, m_And(m_Value(A), m_Value(B))) &&
5225 match(Op1, m_And(m_Value(C), m_Value(D)))) {
5226 Value *X = 0, *Y = 0, *Z = 0;
5229 X = B; Y = D; Z = A;
5230 } else if (A == D) {
5231 X = B; Y = C; Z = A;
5232 } else if (B == C) {
5233 X = A; Y = D; Z = B;
5234 } else if (B == D) {
5235 X = A; Y = C; Z = B;
5238 if (X) { // Build (X^Y) & Z
5239 Op1 = InsertNewInstBefore(BinaryOperator::createXor(X, Y, "tmp"), I);
5240 Op1 = InsertNewInstBefore(BinaryOperator::createAnd(Op1, Z, "tmp"), I);
5241 I.setOperand(0, Op1);
5242 I.setOperand(1, Constant::getNullValue(Op1->getType()));
5247 return Changed ? &I : 0;
5251 /// FoldICmpDivCst - Fold "icmp pred, ([su]div X, DivRHS), CmpRHS" where DivRHS
5252 /// and CmpRHS are both known to be integer constants.
5253 Instruction *InstCombiner::FoldICmpDivCst(ICmpInst &ICI, BinaryOperator *DivI,
5254 ConstantInt *DivRHS) {
5255 ConstantInt *CmpRHS = cast<ConstantInt>(ICI.getOperand(1));
5256 const APInt &CmpRHSV = CmpRHS->getValue();
5258 // FIXME: If the operand types don't match the type of the divide
5259 // then don't attempt this transform. The code below doesn't have the
5260 // logic to deal with a signed divide and an unsigned compare (and
5261 // vice versa). This is because (x /s C1) <s C2 produces different
5262 // results than (x /s C1) <u C2 or (x /u C1) <s C2 or even
5263 // (x /u C1) <u C2. Simply casting the operands and result won't
5264 // work. :( The if statement below tests that condition and bails
5266 bool DivIsSigned = DivI->getOpcode() == Instruction::SDiv;
5267 if (!ICI.isEquality() && DivIsSigned != ICI.isSignedPredicate())
5269 if (DivRHS->isZero())
5270 return 0; // The ProdOV computation fails on divide by zero.
5272 // Compute Prod = CI * DivRHS. We are essentially solving an equation
5273 // of form X/C1=C2. We solve for X by multiplying C1 (DivRHS) and
5274 // C2 (CI). By solving for X we can turn this into a range check
5275 // instead of computing a divide.
5276 ConstantInt *Prod = Multiply(CmpRHS, DivRHS);
5278 // Determine if the product overflows by seeing if the product is
5279 // not equal to the divide. Make sure we do the same kind of divide
5280 // as in the LHS instruction that we're folding.
5281 bool ProdOV = (DivIsSigned ? ConstantExpr::getSDiv(Prod, DivRHS) :
5282 ConstantExpr::getUDiv(Prod, DivRHS)) != CmpRHS;
5284 // Get the ICmp opcode
5285 ICmpInst::Predicate Pred = ICI.getPredicate();
5287 // Figure out the interval that is being checked. For example, a comparison
5288 // like "X /u 5 == 0" is really checking that X is in the interval [0, 5).
5289 // Compute this interval based on the constants involved and the signedness of
5290 // the compare/divide. This computes a half-open interval, keeping track of
5291 // whether either value in the interval overflows. After analysis each
5292 // overflow variable is set to 0 if it's corresponding bound variable is valid
5293 // -1 if overflowed off the bottom end, or +1 if overflowed off the top end.
5294 int LoOverflow = 0, HiOverflow = 0;
5295 ConstantInt *LoBound = 0, *HiBound = 0;
5298 if (!DivIsSigned) { // udiv
5299 // e.g. X/5 op 3 --> [15, 20)
5301 HiOverflow = LoOverflow = ProdOV;
5303 HiOverflow = AddWithOverflow(HiBound, LoBound, DivRHS, false);
5304 } else if (DivRHS->getValue().isStrictlyPositive()) { // Divisor is > 0.
5305 if (CmpRHSV == 0) { // (X / pos) op 0
5306 // Can't overflow. e.g. X/2 op 0 --> [-1, 2)
5307 LoBound = cast<ConstantInt>(ConstantExpr::getNeg(SubOne(DivRHS)));
5309 } else if (CmpRHSV.isStrictlyPositive()) { // (X / pos) op pos
5310 LoBound = Prod; // e.g. X/5 op 3 --> [15, 20)
5311 HiOverflow = LoOverflow = ProdOV;
5313 HiOverflow = AddWithOverflow(HiBound, Prod, DivRHS, true);
5314 } else { // (X / pos) op neg
5315 // e.g. X/5 op -3 --> [-15-4, -15+1) --> [-19, -14)
5316 Constant *DivRHSH = ConstantExpr::getNeg(SubOne(DivRHS));
5317 LoOverflow = AddWithOverflow(LoBound, Prod,
5318 cast<ConstantInt>(DivRHSH), true) ? -1 : 0;
5319 HiBound = AddOne(Prod);
5320 HiOverflow = ProdOV ? -1 : 0;
5322 } else if (DivRHS->getValue().isNegative()) { // Divisor is < 0.
5323 if (CmpRHSV == 0) { // (X / neg) op 0
5324 // e.g. X/-5 op 0 --> [-4, 5)
5325 LoBound = AddOne(DivRHS);
5326 HiBound = cast<ConstantInt>(ConstantExpr::getNeg(DivRHS));
5327 if (HiBound == DivRHS) { // -INTMIN = INTMIN
5328 HiOverflow = 1; // [INTMIN+1, overflow)
5329 HiBound = 0; // e.g. X/INTMIN = 0 --> X > INTMIN
5331 } else if (CmpRHSV.isStrictlyPositive()) { // (X / neg) op pos
5332 // e.g. X/-5 op 3 --> [-19, -14)
5333 HiOverflow = LoOverflow = ProdOV ? -1 : 0;
5335 LoOverflow = AddWithOverflow(LoBound, Prod, AddOne(DivRHS), true) ?-1:0;
5336 HiBound = AddOne(Prod);
5337 } else { // (X / neg) op neg
5338 // e.g. X/-5 op -3 --> [15, 20)
5340 LoOverflow = HiOverflow = ProdOV ? 1 : 0;
5341 HiBound = Subtract(Prod, DivRHS);
5344 // Dividing by a negative swaps the condition. LT <-> GT
5345 Pred = ICmpInst::getSwappedPredicate(Pred);
5348 Value *X = DivI->getOperand(0);
5350 default: assert(0 && "Unhandled icmp opcode!");
5351 case ICmpInst::ICMP_EQ:
5352 if (LoOverflow && HiOverflow)
5353 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
5354 else if (HiOverflow)
5355 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
5356 ICmpInst::ICMP_UGE, X, LoBound);
5357 else if (LoOverflow)
5358 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
5359 ICmpInst::ICMP_ULT, X, HiBound);
5361 return InsertRangeTest(X, LoBound, HiBound, DivIsSigned, true, ICI);
5362 case ICmpInst::ICMP_NE:
5363 if (LoOverflow && HiOverflow)
5364 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
5365 else if (HiOverflow)
5366 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
5367 ICmpInst::ICMP_ULT, X, LoBound);
5368 else if (LoOverflow)
5369 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
5370 ICmpInst::ICMP_UGE, X, HiBound);
5372 return InsertRangeTest(X, LoBound, HiBound, DivIsSigned, false, ICI);
5373 case ICmpInst::ICMP_ULT:
5374 case ICmpInst::ICMP_SLT:
5375 if (LoOverflow == +1) // Low bound is greater than input range.
5376 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
5377 if (LoOverflow == -1) // Low bound is less than input range.
5378 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
5379 return new ICmpInst(Pred, X, LoBound);
5380 case ICmpInst::ICMP_UGT:
5381 case ICmpInst::ICMP_SGT:
5382 if (HiOverflow == +1) // High bound greater than input range.
5383 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
5384 else if (HiOverflow == -1) // High bound less than input range.
5385 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
5386 if (Pred == ICmpInst::ICMP_UGT)
5387 return new ICmpInst(ICmpInst::ICMP_UGE, X, HiBound);
5389 return new ICmpInst(ICmpInst::ICMP_SGE, X, HiBound);
5394 /// visitICmpInstWithInstAndIntCst - Handle "icmp (instr, intcst)".
5396 Instruction *InstCombiner::visitICmpInstWithInstAndIntCst(ICmpInst &ICI,
5399 const APInt &RHSV = RHS->getValue();
5401 switch (LHSI->getOpcode()) {
5402 case Instruction::Xor: // (icmp pred (xor X, XorCST), CI)
5403 if (ConstantInt *XorCST = dyn_cast<ConstantInt>(LHSI->getOperand(1))) {
5404 // If this is a comparison that tests the signbit (X < 0) or (x > -1),
5406 if (ICI.getPredicate() == ICmpInst::ICMP_SLT && RHSV == 0 ||
5407 ICI.getPredicate() == ICmpInst::ICMP_SGT && RHSV.isAllOnesValue()) {
5408 Value *CompareVal = LHSI->getOperand(0);
5410 // If the sign bit of the XorCST is not set, there is no change to
5411 // the operation, just stop using the Xor.
5412 if (!XorCST->getValue().isNegative()) {
5413 ICI.setOperand(0, CompareVal);
5414 AddToWorkList(LHSI);
5418 // Was the old condition true if the operand is positive?
5419 bool isTrueIfPositive = ICI.getPredicate() == ICmpInst::ICMP_SGT;
5421 // If so, the new one isn't.
5422 isTrueIfPositive ^= true;
5424 if (isTrueIfPositive)
5425 return new ICmpInst(ICmpInst::ICMP_SGT, CompareVal, SubOne(RHS));
5427 return new ICmpInst(ICmpInst::ICMP_SLT, CompareVal, AddOne(RHS));
5431 case Instruction::And: // (icmp pred (and X, AndCST), RHS)
5432 if (LHSI->hasOneUse() && isa<ConstantInt>(LHSI->getOperand(1)) &&
5433 LHSI->getOperand(0)->hasOneUse()) {
5434 ConstantInt *AndCST = cast<ConstantInt>(LHSI->getOperand(1));
5436 // If the LHS is an AND of a truncating cast, we can widen the
5437 // and/compare to be the input width without changing the value
5438 // produced, eliminating a cast.
5439 if (TruncInst *Cast = dyn_cast<TruncInst>(LHSI->getOperand(0))) {
5440 // We can do this transformation if either the AND constant does not
5441 // have its sign bit set or if it is an equality comparison.
5442 // Extending a relational comparison when we're checking the sign
5443 // bit would not work.
5444 if (Cast->hasOneUse() &&
5445 (ICI.isEquality() || AndCST->getValue().isNonNegative() &&
5446 RHSV.isNonNegative())) {
5448 cast<IntegerType>(Cast->getOperand(0)->getType())->getBitWidth();
5449 APInt NewCST = AndCST->getValue();
5450 NewCST.zext(BitWidth);
5452 NewCI.zext(BitWidth);
5453 Instruction *NewAnd =
5454 BinaryOperator::createAnd(Cast->getOperand(0),
5455 ConstantInt::get(NewCST),LHSI->getName());
5456 InsertNewInstBefore(NewAnd, ICI);
5457 return new ICmpInst(ICI.getPredicate(), NewAnd,
5458 ConstantInt::get(NewCI));
5462 // If this is: (X >> C1) & C2 != C3 (where any shift and any compare
5463 // could exist), turn it into (X & (C2 << C1)) != (C3 << C1). This
5464 // happens a LOT in code produced by the C front-end, for bitfield
5466 BinaryOperator *Shift = dyn_cast<BinaryOperator>(LHSI->getOperand(0));
5467 if (Shift && !Shift->isShift())
5471 ShAmt = Shift ? dyn_cast<ConstantInt>(Shift->getOperand(1)) : 0;
5472 const Type *Ty = Shift ? Shift->getType() : 0; // Type of the shift.
5473 const Type *AndTy = AndCST->getType(); // Type of the and.
5475 // We can fold this as long as we can't shift unknown bits
5476 // into the mask. This can only happen with signed shift
5477 // rights, as they sign-extend.
5479 bool CanFold = Shift->isLogicalShift();
5481 // To test for the bad case of the signed shr, see if any
5482 // of the bits shifted in could be tested after the mask.
5483 uint32_t TyBits = Ty->getPrimitiveSizeInBits();
5484 int ShAmtVal = TyBits - ShAmt->getLimitedValue(TyBits);
5486 uint32_t BitWidth = AndTy->getPrimitiveSizeInBits();
5487 if ((APInt::getHighBitsSet(BitWidth, BitWidth-ShAmtVal) &
5488 AndCST->getValue()) == 0)
5494 if (Shift->getOpcode() == Instruction::Shl)
5495 NewCst = ConstantExpr::getLShr(RHS, ShAmt);
5497 NewCst = ConstantExpr::getShl(RHS, ShAmt);
5499 // Check to see if we are shifting out any of the bits being
5501 if (ConstantExpr::get(Shift->getOpcode(), NewCst, ShAmt) != RHS) {
5502 // If we shifted bits out, the fold is not going to work out.
5503 // As a special case, check to see if this means that the
5504 // result is always true or false now.
5505 if (ICI.getPredicate() == ICmpInst::ICMP_EQ)
5506 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
5507 if (ICI.getPredicate() == ICmpInst::ICMP_NE)
5508 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
5510 ICI.setOperand(1, NewCst);
5511 Constant *NewAndCST;
5512 if (Shift->getOpcode() == Instruction::Shl)
5513 NewAndCST = ConstantExpr::getLShr(AndCST, ShAmt);
5515 NewAndCST = ConstantExpr::getShl(AndCST, ShAmt);
5516 LHSI->setOperand(1, NewAndCST);
5517 LHSI->setOperand(0, Shift->getOperand(0));
5518 AddToWorkList(Shift); // Shift is dead.
5519 AddUsesToWorkList(ICI);
5525 // Turn ((X >> Y) & C) == 0 into (X & (C << Y)) == 0. The later is
5526 // preferable because it allows the C<<Y expression to be hoisted out
5527 // of a loop if Y is invariant and X is not.
5528 if (Shift && Shift->hasOneUse() && RHSV == 0 &&
5529 ICI.isEquality() && !Shift->isArithmeticShift() &&
5530 isa<Instruction>(Shift->getOperand(0))) {
5533 if (Shift->getOpcode() == Instruction::LShr) {
5534 NS = BinaryOperator::createShl(AndCST,
5535 Shift->getOperand(1), "tmp");
5537 // Insert a logical shift.
5538 NS = BinaryOperator::createLShr(AndCST,
5539 Shift->getOperand(1), "tmp");
5541 InsertNewInstBefore(cast<Instruction>(NS), ICI);
5543 // Compute X & (C << Y).
5544 Instruction *NewAnd =
5545 BinaryOperator::createAnd(Shift->getOperand(0), NS, LHSI->getName());
5546 InsertNewInstBefore(NewAnd, ICI);
5548 ICI.setOperand(0, NewAnd);
5554 case Instruction::Shl: { // (icmp pred (shl X, ShAmt), CI)
5555 ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1));
5558 uint32_t TypeBits = RHSV.getBitWidth();
5560 // Check that the shift amount is in range. If not, don't perform
5561 // undefined shifts. When the shift is visited it will be
5563 if (ShAmt->uge(TypeBits))
5566 if (ICI.isEquality()) {
5567 // If we are comparing against bits always shifted out, the
5568 // comparison cannot succeed.
5570 ConstantExpr::getShl(ConstantExpr::getLShr(RHS, ShAmt), ShAmt);
5571 if (Comp != RHS) {// Comparing against a bit that we know is zero.
5572 bool IsICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
5573 Constant *Cst = ConstantInt::get(Type::Int1Ty, IsICMP_NE);
5574 return ReplaceInstUsesWith(ICI, Cst);
5577 if (LHSI->hasOneUse()) {
5578 // Otherwise strength reduce the shift into an and.
5579 uint32_t ShAmtVal = (uint32_t)ShAmt->getLimitedValue(TypeBits);
5581 ConstantInt::get(APInt::getLowBitsSet(TypeBits, TypeBits-ShAmtVal));
5584 BinaryOperator::createAnd(LHSI->getOperand(0),
5585 Mask, LHSI->getName()+".mask");
5586 Value *And = InsertNewInstBefore(AndI, ICI);
5587 return new ICmpInst(ICI.getPredicate(), And,
5588 ConstantInt::get(RHSV.lshr(ShAmtVal)));
5592 // Otherwise, if this is a comparison of the sign bit, simplify to and/test.
5593 bool TrueIfSigned = false;
5594 if (LHSI->hasOneUse() &&
5595 isSignBitCheck(ICI.getPredicate(), RHS, TrueIfSigned)) {
5596 // (X << 31) <s 0 --> (X&1) != 0
5597 Constant *Mask = ConstantInt::get(APInt(TypeBits, 1) <<
5598 (TypeBits-ShAmt->getZExtValue()-1));
5600 BinaryOperator::createAnd(LHSI->getOperand(0),
5601 Mask, LHSI->getName()+".mask");
5602 Value *And = InsertNewInstBefore(AndI, ICI);
5604 return new ICmpInst(TrueIfSigned ? ICmpInst::ICMP_NE : ICmpInst::ICMP_EQ,
5605 And, Constant::getNullValue(And->getType()));
5610 case Instruction::LShr: // (icmp pred (shr X, ShAmt), CI)
5611 case Instruction::AShr: {
5612 ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1));
5615 if (ICI.isEquality()) {
5616 // Check that the shift amount is in range. If not, don't perform
5617 // undefined shifts. When the shift is visited it will be
5619 uint32_t TypeBits = RHSV.getBitWidth();
5620 if (ShAmt->uge(TypeBits))
5622 uint32_t ShAmtVal = (uint32_t)ShAmt->getLimitedValue(TypeBits);
5624 // If we are comparing against bits always shifted out, the
5625 // comparison cannot succeed.
5626 APInt Comp = RHSV << ShAmtVal;
5627 if (LHSI->getOpcode() == Instruction::LShr)
5628 Comp = Comp.lshr(ShAmtVal);
5630 Comp = Comp.ashr(ShAmtVal);
5632 if (Comp != RHSV) { // Comparing against a bit that we know is zero.
5633 bool IsICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
5634 Constant *Cst = ConstantInt::get(Type::Int1Ty, IsICMP_NE);
5635 return ReplaceInstUsesWith(ICI, Cst);
5638 if (LHSI->hasOneUse() || RHSV == 0) {
5639 // Otherwise strength reduce the shift into an and.
5640 APInt Val(APInt::getHighBitsSet(TypeBits, TypeBits - ShAmtVal));
5641 Constant *Mask = ConstantInt::get(Val);
5644 BinaryOperator::createAnd(LHSI->getOperand(0),
5645 Mask, LHSI->getName()+".mask");
5646 Value *And = InsertNewInstBefore(AndI, ICI);
5647 return new ICmpInst(ICI.getPredicate(), And,
5648 ConstantExpr::getShl(RHS, ShAmt));
5654 case Instruction::SDiv:
5655 case Instruction::UDiv:
5656 // Fold: icmp pred ([us]div X, C1), C2 -> range test
5657 // Fold this div into the comparison, producing a range check.
5658 // Determine, based on the divide type, what the range is being
5659 // checked. If there is an overflow on the low or high side, remember
5660 // it, otherwise compute the range [low, hi) bounding the new value.
5661 // See: InsertRangeTest above for the kinds of replacements possible.
5662 if (ConstantInt *DivRHS = dyn_cast<ConstantInt>(LHSI->getOperand(1)))
5663 if (Instruction *R = FoldICmpDivCst(ICI, cast<BinaryOperator>(LHSI),
5668 case Instruction::Add:
5669 // Fold: icmp pred (add, X, C1), C2
5671 if (!ICI.isEquality()) {
5672 ConstantInt *LHSC = dyn_cast<ConstantInt>(LHSI->getOperand(1));
5674 const APInt &LHSV = LHSC->getValue();
5676 ConstantRange CR = ICI.makeConstantRange(ICI.getPredicate(), RHSV)
5679 if (ICI.isSignedPredicate()) {
5680 if (CR.getLower().isSignBit()) {
5681 return new ICmpInst(ICmpInst::ICMP_SLT, LHSI->getOperand(0),
5682 ConstantInt::get(CR.getUpper()));
5683 } else if (CR.getUpper().isSignBit()) {
5684 return new ICmpInst(ICmpInst::ICMP_SGE, LHSI->getOperand(0),
5685 ConstantInt::get(CR.getLower()));
5688 if (CR.getLower().isMinValue()) {
5689 return new ICmpInst(ICmpInst::ICMP_ULT, LHSI->getOperand(0),
5690 ConstantInt::get(CR.getUpper()));
5691 } else if (CR.getUpper().isMinValue()) {
5692 return new ICmpInst(ICmpInst::ICMP_UGE, LHSI->getOperand(0),
5693 ConstantInt::get(CR.getLower()));
5700 // Simplify icmp_eq and icmp_ne instructions with integer constant RHS.
5701 if (ICI.isEquality()) {
5702 bool isICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
5704 // If the first operand is (add|sub|and|or|xor|rem) with a constant, and
5705 // the second operand is a constant, simplify a bit.
5706 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(LHSI)) {
5707 switch (BO->getOpcode()) {
5708 case Instruction::SRem:
5709 // If we have a signed (X % (2^c)) == 0, turn it into an unsigned one.
5710 if (RHSV == 0 && isa<ConstantInt>(BO->getOperand(1)) &&BO->hasOneUse()){
5711 const APInt &V = cast<ConstantInt>(BO->getOperand(1))->getValue();
5712 if (V.sgt(APInt(V.getBitWidth(), 1)) && V.isPowerOf2()) {
5713 Instruction *NewRem =
5714 BinaryOperator::createURem(BO->getOperand(0), BO->getOperand(1),
5716 InsertNewInstBefore(NewRem, ICI);
5717 return new ICmpInst(ICI.getPredicate(), NewRem,
5718 Constant::getNullValue(BO->getType()));
5722 case Instruction::Add:
5723 // Replace ((add A, B) != C) with (A != C-B) if B & C are constants.
5724 if (ConstantInt *BOp1C = dyn_cast<ConstantInt>(BO->getOperand(1))) {
5725 if (BO->hasOneUse())
5726 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
5727 Subtract(RHS, BOp1C));
5728 } else if (RHSV == 0) {
5729 // Replace ((add A, B) != 0) with (A != -B) if A or B is
5730 // efficiently invertible, or if the add has just this one use.
5731 Value *BOp0 = BO->getOperand(0), *BOp1 = BO->getOperand(1);
5733 if (Value *NegVal = dyn_castNegVal(BOp1))
5734 return new ICmpInst(ICI.getPredicate(), BOp0, NegVal);
5735 else if (Value *NegVal = dyn_castNegVal(BOp0))
5736 return new ICmpInst(ICI.getPredicate(), NegVal, BOp1);
5737 else if (BO->hasOneUse()) {
5738 Instruction *Neg = BinaryOperator::createNeg(BOp1);
5739 InsertNewInstBefore(Neg, ICI);
5741 return new ICmpInst(ICI.getPredicate(), BOp0, Neg);
5745 case Instruction::Xor:
5746 // For the xor case, we can xor two constants together, eliminating
5747 // the explicit xor.
5748 if (Constant *BOC = dyn_cast<Constant>(BO->getOperand(1)))
5749 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
5750 ConstantExpr::getXor(RHS, BOC));
5753 case Instruction::Sub:
5754 // Replace (([sub|xor] A, B) != 0) with (A != B)
5756 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
5760 case Instruction::Or:
5761 // If bits are being or'd in that are not present in the constant we
5762 // are comparing against, then the comparison could never succeed!
5763 if (Constant *BOC = dyn_cast<Constant>(BO->getOperand(1))) {
5764 Constant *NotCI = ConstantExpr::getNot(RHS);
5765 if (!ConstantExpr::getAnd(BOC, NotCI)->isNullValue())
5766 return ReplaceInstUsesWith(ICI, ConstantInt::get(Type::Int1Ty,
5771 case Instruction::And:
5772 if (ConstantInt *BOC = dyn_cast<ConstantInt>(BO->getOperand(1))) {
5773 // If bits are being compared against that are and'd out, then the
5774 // comparison can never succeed!
5775 if ((RHSV & ~BOC->getValue()) != 0)
5776 return ReplaceInstUsesWith(ICI, ConstantInt::get(Type::Int1Ty,
5779 // If we have ((X & C) == C), turn it into ((X & C) != 0).
5780 if (RHS == BOC && RHSV.isPowerOf2())
5781 return new ICmpInst(isICMP_NE ? ICmpInst::ICMP_EQ :
5782 ICmpInst::ICMP_NE, LHSI,
5783 Constant::getNullValue(RHS->getType()));
5785 // Replace (and X, (1 << size(X)-1) != 0) with x s< 0
5786 if (isSignBit(BOC)) {
5787 Value *X = BO->getOperand(0);
5788 Constant *Zero = Constant::getNullValue(X->getType());
5789 ICmpInst::Predicate pred = isICMP_NE ?
5790 ICmpInst::ICMP_SLT : ICmpInst::ICMP_SGE;
5791 return new ICmpInst(pred, X, Zero);
5794 // ((X & ~7) == 0) --> X < 8
5795 if (RHSV == 0 && isHighOnes(BOC)) {
5796 Value *X = BO->getOperand(0);
5797 Constant *NegX = ConstantExpr::getNeg(BOC);
5798 ICmpInst::Predicate pred = isICMP_NE ?
5799 ICmpInst::ICMP_UGE : ICmpInst::ICMP_ULT;
5800 return new ICmpInst(pred, X, NegX);
5805 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(LHSI)) {
5806 // Handle icmp {eq|ne} <intrinsic>, intcst.
5807 if (II->getIntrinsicID() == Intrinsic::bswap) {
5809 ICI.setOperand(0, II->getOperand(1));
5810 ICI.setOperand(1, ConstantInt::get(RHSV.byteSwap()));
5814 } else { // Not a ICMP_EQ/ICMP_NE
5815 // If the LHS is a cast from an integral value of the same size,
5816 // then since we know the RHS is a constant, try to simlify.
5817 if (CastInst *Cast = dyn_cast<CastInst>(LHSI)) {
5818 Value *CastOp = Cast->getOperand(0);
5819 const Type *SrcTy = CastOp->getType();
5820 uint32_t SrcTySize = SrcTy->getPrimitiveSizeInBits();
5821 if (SrcTy->isInteger() &&
5822 SrcTySize == Cast->getType()->getPrimitiveSizeInBits()) {
5823 // If this is an unsigned comparison, try to make the comparison use
5824 // smaller constant values.
5825 if (ICI.getPredicate() == ICmpInst::ICMP_ULT && RHSV.isSignBit()) {
5826 // X u< 128 => X s> -1
5827 return new ICmpInst(ICmpInst::ICMP_SGT, CastOp,
5828 ConstantInt::get(APInt::getAllOnesValue(SrcTySize)));
5829 } else if (ICI.getPredicate() == ICmpInst::ICMP_UGT &&
5830 RHSV == APInt::getSignedMaxValue(SrcTySize)) {
5831 // X u> 127 => X s< 0
5832 return new ICmpInst(ICmpInst::ICMP_SLT, CastOp,
5833 Constant::getNullValue(SrcTy));
5841 /// visitICmpInstWithCastAndCast - Handle icmp (cast x to y), (cast/cst).
5842 /// We only handle extending casts so far.
5844 Instruction *InstCombiner::visitICmpInstWithCastAndCast(ICmpInst &ICI) {
5845 const CastInst *LHSCI = cast<CastInst>(ICI.getOperand(0));
5846 Value *LHSCIOp = LHSCI->getOperand(0);
5847 const Type *SrcTy = LHSCIOp->getType();
5848 const Type *DestTy = LHSCI->getType();
5851 // Turn icmp (ptrtoint x), (ptrtoint/c) into a compare of the input if the
5852 // integer type is the same size as the pointer type.
5853 if (LHSCI->getOpcode() == Instruction::PtrToInt &&
5854 getTargetData().getPointerSizeInBits() ==
5855 cast<IntegerType>(DestTy)->getBitWidth()) {
5857 if (Constant *RHSC = dyn_cast<Constant>(ICI.getOperand(1))) {
5858 RHSOp = ConstantExpr::getIntToPtr(RHSC, SrcTy);
5859 } else if (PtrToIntInst *RHSC = dyn_cast<PtrToIntInst>(ICI.getOperand(1))) {
5860 RHSOp = RHSC->getOperand(0);
5861 // If the pointer types don't match, insert a bitcast.
5862 if (LHSCIOp->getType() != RHSOp->getType())
5863 RHSOp = InsertBitCastBefore(RHSOp, LHSCIOp->getType(), ICI);
5867 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSOp);
5870 // The code below only handles extension cast instructions, so far.
5872 if (LHSCI->getOpcode() != Instruction::ZExt &&
5873 LHSCI->getOpcode() != Instruction::SExt)
5876 bool isSignedExt = LHSCI->getOpcode() == Instruction::SExt;
5877 bool isSignedCmp = ICI.isSignedPredicate();
5879 if (CastInst *CI = dyn_cast<CastInst>(ICI.getOperand(1))) {
5880 // Not an extension from the same type?
5881 RHSCIOp = CI->getOperand(0);
5882 if (RHSCIOp->getType() != LHSCIOp->getType())
5885 // If the signedness of the two casts doesn't agree (i.e. one is a sext
5886 // and the other is a zext), then we can't handle this.
5887 if (CI->getOpcode() != LHSCI->getOpcode())
5890 // Deal with equality cases early.
5891 if (ICI.isEquality())
5892 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSCIOp);
5894 // A signed comparison of sign extended values simplifies into a
5895 // signed comparison.
5896 if (isSignedCmp && isSignedExt)
5897 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSCIOp);
5899 // The other three cases all fold into an unsigned comparison.
5900 return new ICmpInst(ICI.getUnsignedPredicate(), LHSCIOp, RHSCIOp);
5903 // If we aren't dealing with a constant on the RHS, exit early
5904 ConstantInt *CI = dyn_cast<ConstantInt>(ICI.getOperand(1));
5908 // Compute the constant that would happen if we truncated to SrcTy then
5909 // reextended to DestTy.
5910 Constant *Res1 = ConstantExpr::getTrunc(CI, SrcTy);
5911 Constant *Res2 = ConstantExpr::getCast(LHSCI->getOpcode(), Res1, DestTy);
5913 // If the re-extended constant didn't change...
5915 // Make sure that sign of the Cmp and the sign of the Cast are the same.
5916 // For example, we might have:
5917 // %A = sext short %X to uint
5918 // %B = icmp ugt uint %A, 1330
5919 // It is incorrect to transform this into
5920 // %B = icmp ugt short %X, 1330
5921 // because %A may have negative value.
5923 // However, it is OK if SrcTy is bool (See cast-set.ll testcase)
5924 // OR operation is EQ/NE.
5925 if (isSignedExt == isSignedCmp || SrcTy == Type::Int1Ty || ICI.isEquality())
5926 return new ICmpInst(ICI.getPredicate(), LHSCIOp, Res1);
5931 // The re-extended constant changed so the constant cannot be represented
5932 // in the shorter type. Consequently, we cannot emit a simple comparison.
5934 // First, handle some easy cases. We know the result cannot be equal at this
5935 // point so handle the ICI.isEquality() cases
5936 if (ICI.getPredicate() == ICmpInst::ICMP_EQ)
5937 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
5938 if (ICI.getPredicate() == ICmpInst::ICMP_NE)
5939 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
5941 // Evaluate the comparison for LT (we invert for GT below). LE and GE cases
5942 // should have been folded away previously and not enter in here.
5945 // We're performing a signed comparison.
5946 if (cast<ConstantInt>(CI)->getValue().isNegative())
5947 Result = ConstantInt::getFalse(); // X < (small) --> false
5949 Result = ConstantInt::getTrue(); // X < (large) --> true
5951 // We're performing an unsigned comparison.
5953 // We're performing an unsigned comp with a sign extended value.
5954 // This is true if the input is >= 0. [aka >s -1]
5955 Constant *NegOne = ConstantInt::getAllOnesValue(SrcTy);
5956 Result = InsertNewInstBefore(new ICmpInst(ICmpInst::ICMP_SGT, LHSCIOp,
5957 NegOne, ICI.getName()), ICI);
5959 // Unsigned extend & unsigned compare -> always true.
5960 Result = ConstantInt::getTrue();
5964 // Finally, return the value computed.
5965 if (ICI.getPredicate() == ICmpInst::ICMP_ULT ||
5966 ICI.getPredicate() == ICmpInst::ICMP_SLT) {
5967 return ReplaceInstUsesWith(ICI, Result);
5969 assert((ICI.getPredicate()==ICmpInst::ICMP_UGT ||
5970 ICI.getPredicate()==ICmpInst::ICMP_SGT) &&
5971 "ICmp should be folded!");
5972 if (Constant *CI = dyn_cast<Constant>(Result))
5973 return ReplaceInstUsesWith(ICI, ConstantExpr::getNot(CI));
5975 return BinaryOperator::createNot(Result);
5979 Instruction *InstCombiner::visitShl(BinaryOperator &I) {
5980 return commonShiftTransforms(I);
5983 Instruction *InstCombiner::visitLShr(BinaryOperator &I) {
5984 return commonShiftTransforms(I);
5987 Instruction *InstCombiner::visitAShr(BinaryOperator &I) {
5988 if (Instruction *R = commonShiftTransforms(I))
5991 Value *Op0 = I.getOperand(0);
5993 // ashr int -1, X = -1 (for any arithmetic shift rights of ~0)
5994 if (ConstantInt *CSI = dyn_cast<ConstantInt>(Op0))
5995 if (CSI->isAllOnesValue())
5996 return ReplaceInstUsesWith(I, CSI);
5998 // See if we can turn a signed shr into an unsigned shr.
5999 if (MaskedValueIsZero(Op0,
6000 APInt::getSignBit(I.getType()->getPrimitiveSizeInBits())))
6001 return BinaryOperator::createLShr(Op0, I.getOperand(1));
6006 Instruction *InstCombiner::commonShiftTransforms(BinaryOperator &I) {
6007 assert(I.getOperand(1)->getType() == I.getOperand(0)->getType());
6008 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
6010 // shl X, 0 == X and shr X, 0 == X
6011 // shl 0, X == 0 and shr 0, X == 0
6012 if (Op1 == Constant::getNullValue(Op1->getType()) ||
6013 Op0 == Constant::getNullValue(Op0->getType()))
6014 return ReplaceInstUsesWith(I, Op0);
6016 if (isa<UndefValue>(Op0)) {
6017 if (I.getOpcode() == Instruction::AShr) // undef >>s X -> undef
6018 return ReplaceInstUsesWith(I, Op0);
6019 else // undef << X -> 0, undef >>u X -> 0
6020 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
6022 if (isa<UndefValue>(Op1)) {
6023 if (I.getOpcode() == Instruction::AShr) // X >>s undef -> X
6024 return ReplaceInstUsesWith(I, Op0);
6025 else // X << undef, X >>u undef -> 0
6026 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
6029 // Try to fold constant and into select arguments.
6030 if (isa<Constant>(Op0))
6031 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
6032 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
6035 if (ConstantInt *CUI = dyn_cast<ConstantInt>(Op1))
6036 if (Instruction *Res = FoldShiftByConstant(Op0, CUI, I))
6041 Instruction *InstCombiner::FoldShiftByConstant(Value *Op0, ConstantInt *Op1,
6042 BinaryOperator &I) {
6043 bool isLeftShift = I.getOpcode() == Instruction::Shl;
6045 // See if we can simplify any instructions used by the instruction whose sole
6046 // purpose is to compute bits we don't care about.
6047 uint32_t TypeBits = Op0->getType()->getPrimitiveSizeInBits();
6048 APInt KnownZero(TypeBits, 0), KnownOne(TypeBits, 0);
6049 if (SimplifyDemandedBits(&I, APInt::getAllOnesValue(TypeBits),
6050 KnownZero, KnownOne))
6053 // shl uint X, 32 = 0 and shr ubyte Y, 9 = 0, ... just don't eliminate shr
6054 // of a signed value.
6056 if (Op1->uge(TypeBits)) {
6057 if (I.getOpcode() != Instruction::AShr)
6058 return ReplaceInstUsesWith(I, Constant::getNullValue(Op0->getType()));
6060 I.setOperand(1, ConstantInt::get(I.getType(), TypeBits-1));
6065 // ((X*C1) << C2) == (X * (C1 << C2))
6066 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Op0))
6067 if (BO->getOpcode() == Instruction::Mul && isLeftShift)
6068 if (Constant *BOOp = dyn_cast<Constant>(BO->getOperand(1)))
6069 return BinaryOperator::createMul(BO->getOperand(0),
6070 ConstantExpr::getShl(BOOp, Op1));
6072 // Try to fold constant and into select arguments.
6073 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
6074 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
6076 if (isa<PHINode>(Op0))
6077 if (Instruction *NV = FoldOpIntoPhi(I))
6080 // Fold shift2(trunc(shift1(x,c1)), c2) -> trunc(shift2(shift1(x,c1),c2))
6081 if (TruncInst *TI = dyn_cast<TruncInst>(Op0)) {
6082 Instruction *TrOp = dyn_cast<Instruction>(TI->getOperand(0));
6083 // If 'shift2' is an ashr, we would have to get the sign bit into a funny
6084 // place. Don't try to do this transformation in this case. Also, we
6085 // require that the input operand is a shift-by-constant so that we have
6086 // confidence that the shifts will get folded together. We could do this
6087 // xform in more cases, but it is unlikely to be profitable.
6088 if (TrOp && I.isLogicalShift() && TrOp->isShift() &&
6089 isa<ConstantInt>(TrOp->getOperand(1))) {
6090 // Okay, we'll do this xform. Make the shift of shift.
6091 Constant *ShAmt = ConstantExpr::getZExt(Op1, TrOp->getType());
6092 Instruction *NSh = BinaryOperator::create(I.getOpcode(), TrOp, ShAmt,
6094 InsertNewInstBefore(NSh, I); // (shift2 (shift1 & 0x00FF), c2)
6096 // For logical shifts, the truncation has the effect of making the high
6097 // part of the register be zeros. Emulate this by inserting an AND to
6098 // clear the top bits as needed. This 'and' will usually be zapped by
6099 // other xforms later if dead.
6100 unsigned SrcSize = TrOp->getType()->getPrimitiveSizeInBits();
6101 unsigned DstSize = TI->getType()->getPrimitiveSizeInBits();
6102 APInt MaskV(APInt::getLowBitsSet(SrcSize, DstSize));
6104 // The mask we constructed says what the trunc would do if occurring
6105 // between the shifts. We want to know the effect *after* the second
6106 // shift. We know that it is a logical shift by a constant, so adjust the
6107 // mask as appropriate.
6108 if (I.getOpcode() == Instruction::Shl)
6109 MaskV <<= Op1->getZExtValue();
6111 assert(I.getOpcode() == Instruction::LShr && "Unknown logical shift");
6112 MaskV = MaskV.lshr(Op1->getZExtValue());
6115 Instruction *And = BinaryOperator::createAnd(NSh, ConstantInt::get(MaskV),
6117 InsertNewInstBefore(And, I); // shift1 & 0x00FF
6119 // Return the value truncated to the interesting size.
6120 return new TruncInst(And, I.getType());
6124 if (Op0->hasOneUse()) {
6125 if (BinaryOperator *Op0BO = dyn_cast<BinaryOperator>(Op0)) {
6126 // Turn ((X >> C) + Y) << C -> (X + (Y << C)) & (~0 << C)
6129 switch (Op0BO->getOpcode()) {
6131 case Instruction::Add:
6132 case Instruction::And:
6133 case Instruction::Or:
6134 case Instruction::Xor: {
6135 // These operators commute.
6136 // Turn (Y + (X >> C)) << C -> (X + (Y << C)) & (~0 << C)
6137 if (isLeftShift && Op0BO->getOperand(1)->hasOneUse() &&
6138 match(Op0BO->getOperand(1),
6139 m_Shr(m_Value(V1), m_ConstantInt(CC))) && CC == Op1) {
6140 Instruction *YS = BinaryOperator::createShl(
6141 Op0BO->getOperand(0), Op1,
6143 InsertNewInstBefore(YS, I); // (Y << C)
6145 BinaryOperator::create(Op0BO->getOpcode(), YS, V1,
6146 Op0BO->getOperand(1)->getName());
6147 InsertNewInstBefore(X, I); // (X + (Y << C))
6148 uint32_t Op1Val = Op1->getLimitedValue(TypeBits);
6149 return BinaryOperator::createAnd(X, ConstantInt::get(
6150 APInt::getHighBitsSet(TypeBits, TypeBits-Op1Val)));
6153 // Turn (Y + ((X >> C) & CC)) << C -> ((X & (CC << C)) + (Y << C))
6154 Value *Op0BOOp1 = Op0BO->getOperand(1);
6155 if (isLeftShift && Op0BOOp1->hasOneUse() &&
6157 m_And(m_Shr(m_Value(V1), m_Value(V2)),m_ConstantInt(CC))) &&
6158 cast<BinaryOperator>(Op0BOOp1)->getOperand(0)->hasOneUse() &&
6160 Instruction *YS = BinaryOperator::createShl(
6161 Op0BO->getOperand(0), Op1,
6163 InsertNewInstBefore(YS, I); // (Y << C)
6165 BinaryOperator::createAnd(V1, ConstantExpr::getShl(CC, Op1),
6166 V1->getName()+".mask");
6167 InsertNewInstBefore(XM, I); // X & (CC << C)
6169 return BinaryOperator::create(Op0BO->getOpcode(), YS, XM);
6174 case Instruction::Sub: {
6175 // Turn ((X >> C) + Y) << C -> (X + (Y << C)) & (~0 << C)
6176 if (isLeftShift && Op0BO->getOperand(0)->hasOneUse() &&
6177 match(Op0BO->getOperand(0),
6178 m_Shr(m_Value(V1), m_ConstantInt(CC))) && CC == Op1) {
6179 Instruction *YS = BinaryOperator::createShl(
6180 Op0BO->getOperand(1), Op1,
6182 InsertNewInstBefore(YS, I); // (Y << C)
6184 BinaryOperator::create(Op0BO->getOpcode(), V1, YS,
6185 Op0BO->getOperand(0)->getName());
6186 InsertNewInstBefore(X, I); // (X + (Y << C))
6187 uint32_t Op1Val = Op1->getLimitedValue(TypeBits);
6188 return BinaryOperator::createAnd(X, ConstantInt::get(
6189 APInt::getHighBitsSet(TypeBits, TypeBits-Op1Val)));
6192 // Turn (((X >> C)&CC) + Y) << C -> (X + (Y << C)) & (CC << C)
6193 if (isLeftShift && Op0BO->getOperand(0)->hasOneUse() &&
6194 match(Op0BO->getOperand(0),
6195 m_And(m_Shr(m_Value(V1), m_Value(V2)),
6196 m_ConstantInt(CC))) && V2 == Op1 &&
6197 cast<BinaryOperator>(Op0BO->getOperand(0))
6198 ->getOperand(0)->hasOneUse()) {
6199 Instruction *YS = BinaryOperator::createShl(
6200 Op0BO->getOperand(1), Op1,
6202 InsertNewInstBefore(YS, I); // (Y << C)
6204 BinaryOperator::createAnd(V1, ConstantExpr::getShl(CC, Op1),
6205 V1->getName()+".mask");
6206 InsertNewInstBefore(XM, I); // X & (CC << C)
6208 return BinaryOperator::create(Op0BO->getOpcode(), XM, YS);
6216 // If the operand is an bitwise operator with a constant RHS, and the
6217 // shift is the only use, we can pull it out of the shift.
6218 if (ConstantInt *Op0C = dyn_cast<ConstantInt>(Op0BO->getOperand(1))) {
6219 bool isValid = true; // Valid only for And, Or, Xor
6220 bool highBitSet = false; // Transform if high bit of constant set?
6222 switch (Op0BO->getOpcode()) {
6223 default: isValid = false; break; // Do not perform transform!
6224 case Instruction::Add:
6225 isValid = isLeftShift;
6227 case Instruction::Or:
6228 case Instruction::Xor:
6231 case Instruction::And:
6236 // If this is a signed shift right, and the high bit is modified
6237 // by the logical operation, do not perform the transformation.
6238 // The highBitSet boolean indicates the value of the high bit of
6239 // the constant which would cause it to be modified for this
6242 if (isValid && I.getOpcode() == Instruction::AShr)
6243 isValid = Op0C->getValue()[TypeBits-1] == highBitSet;
6246 Constant *NewRHS = ConstantExpr::get(I.getOpcode(), Op0C, Op1);
6248 Instruction *NewShift =
6249 BinaryOperator::create(I.getOpcode(), Op0BO->getOperand(0), Op1);
6250 InsertNewInstBefore(NewShift, I);
6251 NewShift->takeName(Op0BO);
6253 return BinaryOperator::create(Op0BO->getOpcode(), NewShift,
6260 // Find out if this is a shift of a shift by a constant.
6261 BinaryOperator *ShiftOp = dyn_cast<BinaryOperator>(Op0);
6262 if (ShiftOp && !ShiftOp->isShift())
6265 if (ShiftOp && isa<ConstantInt>(ShiftOp->getOperand(1))) {
6266 ConstantInt *ShiftAmt1C = cast<ConstantInt>(ShiftOp->getOperand(1));
6267 uint32_t ShiftAmt1 = ShiftAmt1C->getLimitedValue(TypeBits);
6268 uint32_t ShiftAmt2 = Op1->getLimitedValue(TypeBits);
6269 assert(ShiftAmt2 != 0 && "Should have been simplified earlier");
6270 if (ShiftAmt1 == 0) return 0; // Will be simplified in the future.
6271 Value *X = ShiftOp->getOperand(0);
6273 uint32_t AmtSum = ShiftAmt1+ShiftAmt2; // Fold into one big shift.
6274 if (AmtSum > TypeBits)
6277 const IntegerType *Ty = cast<IntegerType>(I.getType());
6279 // Check for (X << c1) << c2 and (X >> c1) >> c2
6280 if (I.getOpcode() == ShiftOp->getOpcode()) {
6281 return BinaryOperator::create(I.getOpcode(), X,
6282 ConstantInt::get(Ty, AmtSum));
6283 } else if (ShiftOp->getOpcode() == Instruction::LShr &&
6284 I.getOpcode() == Instruction::AShr) {
6285 // ((X >>u C1) >>s C2) -> (X >>u (C1+C2)) since C1 != 0.
6286 return BinaryOperator::createLShr(X, ConstantInt::get(Ty, AmtSum));
6287 } else if (ShiftOp->getOpcode() == Instruction::AShr &&
6288 I.getOpcode() == Instruction::LShr) {
6289 // ((X >>s C1) >>u C2) -> ((X >>s (C1+C2)) & mask) since C1 != 0.
6290 Instruction *Shift =
6291 BinaryOperator::createAShr(X, ConstantInt::get(Ty, AmtSum));
6292 InsertNewInstBefore(Shift, I);
6294 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt2));
6295 return BinaryOperator::createAnd(Shift, ConstantInt::get(Mask));
6298 // Okay, if we get here, one shift must be left, and the other shift must be
6299 // right. See if the amounts are equal.
6300 if (ShiftAmt1 == ShiftAmt2) {
6301 // If we have ((X >>? C) << C), turn this into X & (-1 << C).
6302 if (I.getOpcode() == Instruction::Shl) {
6303 APInt Mask(APInt::getHighBitsSet(TypeBits, TypeBits - ShiftAmt1));
6304 return BinaryOperator::createAnd(X, ConstantInt::get(Mask));
6306 // If we have ((X << C) >>u C), turn this into X & (-1 >>u C).
6307 if (I.getOpcode() == Instruction::LShr) {
6308 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt1));
6309 return BinaryOperator::createAnd(X, ConstantInt::get(Mask));
6311 // We can simplify ((X << C) >>s C) into a trunc + sext.
6312 // NOTE: we could do this for any C, but that would make 'unusual' integer
6313 // types. For now, just stick to ones well-supported by the code
6315 const Type *SExtType = 0;
6316 switch (Ty->getBitWidth() - ShiftAmt1) {
6323 SExtType = IntegerType::get(Ty->getBitWidth() - ShiftAmt1);
6328 Instruction *NewTrunc = new TruncInst(X, SExtType, "sext");
6329 InsertNewInstBefore(NewTrunc, I);
6330 return new SExtInst(NewTrunc, Ty);
6332 // Otherwise, we can't handle it yet.
6333 } else if (ShiftAmt1 < ShiftAmt2) {
6334 uint32_t ShiftDiff = ShiftAmt2-ShiftAmt1;
6336 // (X >>? C1) << C2 --> X << (C2-C1) & (-1 << C2)
6337 if (I.getOpcode() == Instruction::Shl) {
6338 assert(ShiftOp->getOpcode() == Instruction::LShr ||
6339 ShiftOp->getOpcode() == Instruction::AShr);
6340 Instruction *Shift =
6341 BinaryOperator::createShl(X, ConstantInt::get(Ty, ShiftDiff));
6342 InsertNewInstBefore(Shift, I);
6344 APInt Mask(APInt::getHighBitsSet(TypeBits, TypeBits - ShiftAmt2));
6345 return BinaryOperator::createAnd(Shift, ConstantInt::get(Mask));
6348 // (X << C1) >>u C2 --> X >>u (C2-C1) & (-1 >> C2)
6349 if (I.getOpcode() == Instruction::LShr) {
6350 assert(ShiftOp->getOpcode() == Instruction::Shl);
6351 Instruction *Shift =
6352 BinaryOperator::createLShr(X, ConstantInt::get(Ty, ShiftDiff));
6353 InsertNewInstBefore(Shift, I);
6355 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt2));
6356 return BinaryOperator::createAnd(Shift, ConstantInt::get(Mask));
6359 // We can't handle (X << C1) >>s C2, it shifts arbitrary bits in.
6361 assert(ShiftAmt2 < ShiftAmt1);
6362 uint32_t ShiftDiff = ShiftAmt1-ShiftAmt2;
6364 // (X >>? C1) << C2 --> X >>? (C1-C2) & (-1 << C2)
6365 if (I.getOpcode() == Instruction::Shl) {
6366 assert(ShiftOp->getOpcode() == Instruction::LShr ||
6367 ShiftOp->getOpcode() == Instruction::AShr);
6368 Instruction *Shift =
6369 BinaryOperator::create(ShiftOp->getOpcode(), X,
6370 ConstantInt::get(Ty, ShiftDiff));
6371 InsertNewInstBefore(Shift, I);
6373 APInt Mask(APInt::getHighBitsSet(TypeBits, TypeBits - ShiftAmt2));
6374 return BinaryOperator::createAnd(Shift, ConstantInt::get(Mask));
6377 // (X << C1) >>u C2 --> X << (C1-C2) & (-1 >> C2)
6378 if (I.getOpcode() == Instruction::LShr) {
6379 assert(ShiftOp->getOpcode() == Instruction::Shl);
6380 Instruction *Shift =
6381 BinaryOperator::createShl(X, ConstantInt::get(Ty, ShiftDiff));
6382 InsertNewInstBefore(Shift, I);
6384 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt2));
6385 return BinaryOperator::createAnd(Shift, ConstantInt::get(Mask));
6388 // We can't handle (X << C1) >>a C2, it shifts arbitrary bits in.
6395 /// DecomposeSimpleLinearExpr - Analyze 'Val', seeing if it is a simple linear
6396 /// expression. If so, decompose it, returning some value X, such that Val is
6399 static Value *DecomposeSimpleLinearExpr(Value *Val, unsigned &Scale,
6401 assert(Val->getType() == Type::Int32Ty && "Unexpected allocation size type!");
6402 if (ConstantInt *CI = dyn_cast<ConstantInt>(Val)) {
6403 Offset = CI->getZExtValue();
6405 return ConstantInt::get(Type::Int32Ty, 0);
6406 } else if (BinaryOperator *I = dyn_cast<BinaryOperator>(Val)) {
6407 if (ConstantInt *RHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
6408 if (I->getOpcode() == Instruction::Shl) {
6409 // This is a value scaled by '1 << the shift amt'.
6410 Scale = 1U << RHS->getZExtValue();
6412 return I->getOperand(0);
6413 } else if (I->getOpcode() == Instruction::Mul) {
6414 // This value is scaled by 'RHS'.
6415 Scale = RHS->getZExtValue();
6417 return I->getOperand(0);
6418 } else if (I->getOpcode() == Instruction::Add) {
6419 // We have X+C. Check to see if we really have (X*C2)+C1,
6420 // where C1 is divisible by C2.
6423 DecomposeSimpleLinearExpr(I->getOperand(0), SubScale, Offset);
6424 Offset += RHS->getZExtValue();
6431 // Otherwise, we can't look past this.
6438 /// PromoteCastOfAllocation - If we find a cast of an allocation instruction,
6439 /// try to eliminate the cast by moving the type information into the alloc.
6440 Instruction *InstCombiner::PromoteCastOfAllocation(BitCastInst &CI,
6441 AllocationInst &AI) {
6442 const PointerType *PTy = cast<PointerType>(CI.getType());
6444 // Remove any uses of AI that are dead.
6445 assert(!CI.use_empty() && "Dead instructions should be removed earlier!");
6447 for (Value::use_iterator UI = AI.use_begin(), E = AI.use_end(); UI != E; ) {
6448 Instruction *User = cast<Instruction>(*UI++);
6449 if (isInstructionTriviallyDead(User)) {
6450 while (UI != E && *UI == User)
6451 ++UI; // If this instruction uses AI more than once, don't break UI.
6454 DOUT << "IC: DCE: " << *User;
6455 EraseInstFromFunction(*User);
6459 // Get the type really allocated and the type casted to.
6460 const Type *AllocElTy = AI.getAllocatedType();
6461 const Type *CastElTy = PTy->getElementType();
6462 if (!AllocElTy->isSized() || !CastElTy->isSized()) return 0;
6464 unsigned AllocElTyAlign = TD->getABITypeAlignment(AllocElTy);
6465 unsigned CastElTyAlign = TD->getABITypeAlignment(CastElTy);
6466 if (CastElTyAlign < AllocElTyAlign) return 0;
6468 // If the allocation has multiple uses, only promote it if we are strictly
6469 // increasing the alignment of the resultant allocation. If we keep it the
6470 // same, we open the door to infinite loops of various kinds.
6471 if (!AI.hasOneUse() && CastElTyAlign == AllocElTyAlign) return 0;
6473 uint64_t AllocElTySize = TD->getABITypeSize(AllocElTy);
6474 uint64_t CastElTySize = TD->getABITypeSize(CastElTy);
6475 if (CastElTySize == 0 || AllocElTySize == 0) return 0;
6477 // See if we can satisfy the modulus by pulling a scale out of the array
6479 unsigned ArraySizeScale;
6481 Value *NumElements = // See if the array size is a decomposable linear expr.
6482 DecomposeSimpleLinearExpr(AI.getOperand(0), ArraySizeScale, ArrayOffset);
6484 // If we can now satisfy the modulus, by using a non-1 scale, we really can
6486 if ((AllocElTySize*ArraySizeScale) % CastElTySize != 0 ||
6487 (AllocElTySize*ArrayOffset ) % CastElTySize != 0) return 0;
6489 unsigned Scale = (AllocElTySize*ArraySizeScale)/CastElTySize;
6494 // If the allocation size is constant, form a constant mul expression
6495 Amt = ConstantInt::get(Type::Int32Ty, Scale);
6496 if (isa<ConstantInt>(NumElements))
6497 Amt = Multiply(cast<ConstantInt>(NumElements), cast<ConstantInt>(Amt));
6498 // otherwise multiply the amount and the number of elements
6499 else if (Scale != 1) {
6500 Instruction *Tmp = BinaryOperator::createMul(Amt, NumElements, "tmp");
6501 Amt = InsertNewInstBefore(Tmp, AI);
6505 if (int Offset = (AllocElTySize*ArrayOffset)/CastElTySize) {
6506 Value *Off = ConstantInt::get(Type::Int32Ty, Offset, true);
6507 Instruction *Tmp = BinaryOperator::createAdd(Amt, Off, "tmp");
6508 Amt = InsertNewInstBefore(Tmp, AI);
6511 AllocationInst *New;
6512 if (isa<MallocInst>(AI))
6513 New = new MallocInst(CastElTy, Amt, AI.getAlignment());
6515 New = new AllocaInst(CastElTy, Amt, AI.getAlignment());
6516 InsertNewInstBefore(New, AI);
6519 // If the allocation has multiple uses, insert a cast and change all things
6520 // that used it to use the new cast. This will also hack on CI, but it will
6522 if (!AI.hasOneUse()) {
6523 AddUsesToWorkList(AI);
6524 // New is the allocation instruction, pointer typed. AI is the original
6525 // allocation instruction, also pointer typed. Thus, cast to use is BitCast.
6526 CastInst *NewCast = new BitCastInst(New, AI.getType(), "tmpcast");
6527 InsertNewInstBefore(NewCast, AI);
6528 AI.replaceAllUsesWith(NewCast);
6530 return ReplaceInstUsesWith(CI, New);
6533 /// CanEvaluateInDifferentType - Return true if we can take the specified value
6534 /// and return it as type Ty without inserting any new casts and without
6535 /// changing the computed value. This is used by code that tries to decide
6536 /// whether promoting or shrinking integer operations to wider or smaller types
6537 /// will allow us to eliminate a truncate or extend.
6539 /// This is a truncation operation if Ty is smaller than V->getType(), or an
6540 /// extension operation if Ty is larger.
6541 static bool CanEvaluateInDifferentType(Value *V, const IntegerType *Ty,
6542 unsigned CastOpc, int &NumCastsRemoved) {
6543 // We can always evaluate constants in another type.
6544 if (isa<ConstantInt>(V))
6547 Instruction *I = dyn_cast<Instruction>(V);
6548 if (!I) return false;
6550 const IntegerType *OrigTy = cast<IntegerType>(V->getType());
6552 // If this is an extension or truncate, we can often eliminate it.
6553 if (isa<TruncInst>(I) || isa<ZExtInst>(I) || isa<SExtInst>(I)) {
6554 // If this is a cast from the destination type, we can trivially eliminate
6555 // it, and this will remove a cast overall.
6556 if (I->getOperand(0)->getType() == Ty) {
6557 // If the first operand is itself a cast, and is eliminable, do not count
6558 // this as an eliminable cast. We would prefer to eliminate those two
6560 if (!isa<CastInst>(I->getOperand(0)))
6566 // We can't extend or shrink something that has multiple uses: doing so would
6567 // require duplicating the instruction in general, which isn't profitable.
6568 if (!I->hasOneUse()) return false;
6570 switch (I->getOpcode()) {
6571 case Instruction::Add:
6572 case Instruction::Sub:
6573 case Instruction::And:
6574 case Instruction::Or:
6575 case Instruction::Xor:
6576 // These operators can all arbitrarily be extended or truncated.
6577 return CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc,
6579 CanEvaluateInDifferentType(I->getOperand(1), Ty, CastOpc,
6582 case Instruction::Mul:
6583 // A multiply can be truncated by truncating its operands.
6584 return Ty->getBitWidth() < OrigTy->getBitWidth() &&
6585 CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc,
6587 CanEvaluateInDifferentType(I->getOperand(1), Ty, CastOpc,
6590 case Instruction::Shl:
6591 // If we are truncating the result of this SHL, and if it's a shift of a
6592 // constant amount, we can always perform a SHL in a smaller type.
6593 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) {
6594 uint32_t BitWidth = Ty->getBitWidth();
6595 if (BitWidth < OrigTy->getBitWidth() &&
6596 CI->getLimitedValue(BitWidth) < BitWidth)
6597 return CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc,
6601 case Instruction::LShr:
6602 // If this is a truncate of a logical shr, we can truncate it to a smaller
6603 // lshr iff we know that the bits we would otherwise be shifting in are
6605 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) {
6606 uint32_t OrigBitWidth = OrigTy->getBitWidth();
6607 uint32_t BitWidth = Ty->getBitWidth();
6608 if (BitWidth < OrigBitWidth &&
6609 MaskedValueIsZero(I->getOperand(0),
6610 APInt::getHighBitsSet(OrigBitWidth, OrigBitWidth-BitWidth)) &&
6611 CI->getLimitedValue(BitWidth) < BitWidth) {
6612 return CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc,
6617 case Instruction::ZExt:
6618 case Instruction::SExt:
6619 case Instruction::Trunc:
6620 // If this is the same kind of case as our original (e.g. zext+zext), we
6621 // can safely replace it. Note that replacing it does not reduce the number
6622 // of casts in the input.
6623 if (I->getOpcode() == CastOpc)
6628 // TODO: Can handle more cases here.
6635 /// EvaluateInDifferentType - Given an expression that
6636 /// CanEvaluateInDifferentType returns true for, actually insert the code to
6637 /// evaluate the expression.
6638 Value *InstCombiner::EvaluateInDifferentType(Value *V, const Type *Ty,
6640 if (Constant *C = dyn_cast<Constant>(V))
6641 return ConstantExpr::getIntegerCast(C, Ty, isSigned /*Sext or ZExt*/);
6643 // Otherwise, it must be an instruction.
6644 Instruction *I = cast<Instruction>(V);
6645 Instruction *Res = 0;
6646 switch (I->getOpcode()) {
6647 case Instruction::Add:
6648 case Instruction::Sub:
6649 case Instruction::Mul:
6650 case Instruction::And:
6651 case Instruction::Or:
6652 case Instruction::Xor:
6653 case Instruction::AShr:
6654 case Instruction::LShr:
6655 case Instruction::Shl: {
6656 Value *LHS = EvaluateInDifferentType(I->getOperand(0), Ty, isSigned);
6657 Value *RHS = EvaluateInDifferentType(I->getOperand(1), Ty, isSigned);
6658 Res = BinaryOperator::create((Instruction::BinaryOps)I->getOpcode(),
6659 LHS, RHS, I->getName());
6662 case Instruction::Trunc:
6663 case Instruction::ZExt:
6664 case Instruction::SExt:
6665 // If the source type of the cast is the type we're trying for then we can
6666 // just return the source. There's no need to insert it because it is not
6668 if (I->getOperand(0)->getType() == Ty)
6669 return I->getOperand(0);
6671 // Otherwise, must be the same type of case, so just reinsert a new one.
6672 Res = CastInst::create(cast<CastInst>(I)->getOpcode(), I->getOperand(0),
6676 // TODO: Can handle more cases here.
6677 assert(0 && "Unreachable!");
6681 return InsertNewInstBefore(Res, *I);
6684 /// @brief Implement the transforms common to all CastInst visitors.
6685 Instruction *InstCombiner::commonCastTransforms(CastInst &CI) {
6686 Value *Src = CI.getOperand(0);
6688 // Many cases of "cast of a cast" are eliminable. If it's eliminable we just
6689 // eliminate it now.
6690 if (CastInst *CSrc = dyn_cast<CastInst>(Src)) { // A->B->C cast
6691 if (Instruction::CastOps opc =
6692 isEliminableCastPair(CSrc, CI.getOpcode(), CI.getType(), TD)) {
6693 // The first cast (CSrc) is eliminable so we need to fix up or replace
6694 // the second cast (CI). CSrc will then have a good chance of being dead.
6695 return CastInst::create(opc, CSrc->getOperand(0), CI.getType());
6699 // If we are casting a select then fold the cast into the select
6700 if (SelectInst *SI = dyn_cast<SelectInst>(Src))
6701 if (Instruction *NV = FoldOpIntoSelect(CI, SI, this))
6704 // If we are casting a PHI then fold the cast into the PHI
6705 if (isa<PHINode>(Src))
6706 if (Instruction *NV = FoldOpIntoPhi(CI))
6712 /// @brief Implement the transforms for cast of pointer (bitcast/ptrtoint)
6713 Instruction *InstCombiner::commonPointerCastTransforms(CastInst &CI) {
6714 Value *Src = CI.getOperand(0);
6716 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Src)) {
6717 // If casting the result of a getelementptr instruction with no offset, turn
6718 // this into a cast of the original pointer!
6719 if (GEP->hasAllZeroIndices()) {
6720 // Changing the cast operand is usually not a good idea but it is safe
6721 // here because the pointer operand is being replaced with another
6722 // pointer operand so the opcode doesn't need to change.
6724 CI.setOperand(0, GEP->getOperand(0));
6728 // If the GEP has a single use, and the base pointer is a bitcast, and the
6729 // GEP computes a constant offset, see if we can convert these three
6730 // instructions into fewer. This typically happens with unions and other
6731 // non-type-safe code.
6732 if (GEP->hasOneUse() && isa<BitCastInst>(GEP->getOperand(0))) {
6733 if (GEP->hasAllConstantIndices()) {
6734 // We are guaranteed to get a constant from EmitGEPOffset.
6735 ConstantInt *OffsetV = cast<ConstantInt>(EmitGEPOffset(GEP, CI, *this));
6736 int64_t Offset = OffsetV->getSExtValue();
6738 // Get the base pointer input of the bitcast, and the type it points to.
6739 Value *OrigBase = cast<BitCastInst>(GEP->getOperand(0))->getOperand(0);
6740 const Type *GEPIdxTy =
6741 cast<PointerType>(OrigBase->getType())->getElementType();
6742 if (GEPIdxTy->isSized()) {
6743 SmallVector<Value*, 8> NewIndices;
6745 // Start with the index over the outer type. Note that the type size
6746 // might be zero (even if the offset isn't zero) if the indexed type
6747 // is something like [0 x {int, int}]
6748 const Type *IntPtrTy = TD->getIntPtrType();
6749 int64_t FirstIdx = 0;
6750 if (int64_t TySize = TD->getABITypeSize(GEPIdxTy)) {
6751 FirstIdx = Offset/TySize;
6754 // Handle silly modulus not returning values values [0..TySize).
6758 assert(Offset >= 0);
6760 assert((uint64_t)Offset < (uint64_t)TySize &&"Out of range offset");
6763 NewIndices.push_back(ConstantInt::get(IntPtrTy, FirstIdx));
6765 // Index into the types. If we fail, set OrigBase to null.
6767 if (const StructType *STy = dyn_cast<StructType>(GEPIdxTy)) {
6768 const StructLayout *SL = TD->getStructLayout(STy);
6769 if (Offset < (int64_t)SL->getSizeInBytes()) {
6770 unsigned Elt = SL->getElementContainingOffset(Offset);
6771 NewIndices.push_back(ConstantInt::get(Type::Int32Ty, Elt));
6773 Offset -= SL->getElementOffset(Elt);
6774 GEPIdxTy = STy->getElementType(Elt);
6776 // Otherwise, we can't index into this, bail out.
6780 } else if (isa<ArrayType>(GEPIdxTy) || isa<VectorType>(GEPIdxTy)) {
6781 const SequentialType *STy = cast<SequentialType>(GEPIdxTy);
6782 if (uint64_t EltSize = TD->getABITypeSize(STy->getElementType())){
6783 NewIndices.push_back(ConstantInt::get(IntPtrTy,Offset/EltSize));
6786 NewIndices.push_back(ConstantInt::get(IntPtrTy, 0));
6788 GEPIdxTy = STy->getElementType();
6790 // Otherwise, we can't index into this, bail out.
6796 // If we were able to index down into an element, create the GEP
6797 // and bitcast the result. This eliminates one bitcast, potentially
6799 Instruction *NGEP = new GetElementPtrInst(OrigBase,
6801 NewIndices.end(), "");
6802 InsertNewInstBefore(NGEP, CI);
6803 NGEP->takeName(GEP);
6805 if (isa<BitCastInst>(CI))
6806 return new BitCastInst(NGEP, CI.getType());
6807 assert(isa<PtrToIntInst>(CI));
6808 return new PtrToIntInst(NGEP, CI.getType());
6815 return commonCastTransforms(CI);
6820 /// Only the TRUNC, ZEXT, SEXT, and BITCAST can both operand and result as
6821 /// integer types. This function implements the common transforms for all those
6823 /// @brief Implement the transforms common to CastInst with integer operands
6824 Instruction *InstCombiner::commonIntCastTransforms(CastInst &CI) {
6825 if (Instruction *Result = commonCastTransforms(CI))
6828 Value *Src = CI.getOperand(0);
6829 const Type *SrcTy = Src->getType();
6830 const Type *DestTy = CI.getType();
6831 uint32_t SrcBitSize = SrcTy->getPrimitiveSizeInBits();
6832 uint32_t DestBitSize = DestTy->getPrimitiveSizeInBits();
6834 // See if we can simplify any instructions used by the LHS whose sole
6835 // purpose is to compute bits we don't care about.
6836 APInt KnownZero(DestBitSize, 0), KnownOne(DestBitSize, 0);
6837 if (SimplifyDemandedBits(&CI, APInt::getAllOnesValue(DestBitSize),
6838 KnownZero, KnownOne))
6841 // If the source isn't an instruction or has more than one use then we
6842 // can't do anything more.
6843 Instruction *SrcI = dyn_cast<Instruction>(Src);
6844 if (!SrcI || !Src->hasOneUse())
6847 // Attempt to propagate the cast into the instruction for int->int casts.
6848 int NumCastsRemoved = 0;
6849 if (!isa<BitCastInst>(CI) &&
6850 CanEvaluateInDifferentType(SrcI, cast<IntegerType>(DestTy),
6851 CI.getOpcode(), NumCastsRemoved)) {
6852 // If this cast is a truncate, evaluting in a different type always
6853 // eliminates the cast, so it is always a win. If this is a zero-extension,
6854 // we need to do an AND to maintain the clear top-part of the computation,
6855 // so we require that the input have eliminated at least one cast. If this
6856 // is a sign extension, we insert two new casts (to do the extension) so we
6857 // require that two casts have been eliminated.
6859 switch (CI.getOpcode()) {
6861 // All the others use floating point so we shouldn't actually
6862 // get here because of the check above.
6863 assert(0 && "Unknown cast type");
6864 case Instruction::Trunc:
6867 case Instruction::ZExt:
6868 DoXForm = NumCastsRemoved >= 1;
6870 case Instruction::SExt:
6871 DoXForm = NumCastsRemoved >= 2;
6876 Value *Res = EvaluateInDifferentType(SrcI, DestTy,
6877 CI.getOpcode() == Instruction::SExt);
6878 assert(Res->getType() == DestTy);
6879 switch (CI.getOpcode()) {
6880 default: assert(0 && "Unknown cast type!");
6881 case Instruction::Trunc:
6882 case Instruction::BitCast:
6883 // Just replace this cast with the result.
6884 return ReplaceInstUsesWith(CI, Res);
6885 case Instruction::ZExt: {
6886 // We need to emit an AND to clear the high bits.
6887 assert(SrcBitSize < DestBitSize && "Not a zext?");
6888 Constant *C = ConstantInt::get(APInt::getLowBitsSet(DestBitSize,
6890 return BinaryOperator::createAnd(Res, C);
6892 case Instruction::SExt:
6893 // We need to emit a cast to truncate, then a cast to sext.
6894 return CastInst::create(Instruction::SExt,
6895 InsertCastBefore(Instruction::Trunc, Res, Src->getType(),
6901 Value *Op0 = SrcI->getNumOperands() > 0 ? SrcI->getOperand(0) : 0;
6902 Value *Op1 = SrcI->getNumOperands() > 1 ? SrcI->getOperand(1) : 0;
6904 switch (SrcI->getOpcode()) {
6905 case Instruction::Add:
6906 case Instruction::Mul:
6907 case Instruction::And:
6908 case Instruction::Or:
6909 case Instruction::Xor:
6910 // If we are discarding information, rewrite.
6911 if (DestBitSize <= SrcBitSize && DestBitSize != 1) {
6912 // Don't insert two casts if they cannot be eliminated. We allow
6913 // two casts to be inserted if the sizes are the same. This could
6914 // only be converting signedness, which is a noop.
6915 if (DestBitSize == SrcBitSize ||
6916 !ValueRequiresCast(CI.getOpcode(), Op1, DestTy,TD) ||
6917 !ValueRequiresCast(CI.getOpcode(), Op0, DestTy, TD)) {
6918 Instruction::CastOps opcode = CI.getOpcode();
6919 Value *Op0c = InsertOperandCastBefore(opcode, Op0, DestTy, SrcI);
6920 Value *Op1c = InsertOperandCastBefore(opcode, Op1, DestTy, SrcI);
6921 return BinaryOperator::create(
6922 cast<BinaryOperator>(SrcI)->getOpcode(), Op0c, Op1c);
6926 // cast (xor bool X, true) to int --> xor (cast bool X to int), 1
6927 if (isa<ZExtInst>(CI) && SrcBitSize == 1 &&
6928 SrcI->getOpcode() == Instruction::Xor &&
6929 Op1 == ConstantInt::getTrue() &&
6930 (!Op0->hasOneUse() || !isa<CmpInst>(Op0))) {
6931 Value *New = InsertOperandCastBefore(Instruction::ZExt, Op0, DestTy, &CI);
6932 return BinaryOperator::createXor(New, ConstantInt::get(CI.getType(), 1));
6935 case Instruction::SDiv:
6936 case Instruction::UDiv:
6937 case Instruction::SRem:
6938 case Instruction::URem:
6939 // If we are just changing the sign, rewrite.
6940 if (DestBitSize == SrcBitSize) {
6941 // Don't insert two casts if they cannot be eliminated. We allow
6942 // two casts to be inserted if the sizes are the same. This could
6943 // only be converting signedness, which is a noop.
6944 if (!ValueRequiresCast(CI.getOpcode(), Op1, DestTy, TD) ||
6945 !ValueRequiresCast(CI.getOpcode(), Op0, DestTy, TD)) {
6946 Value *Op0c = InsertOperandCastBefore(Instruction::BitCast,
6948 Value *Op1c = InsertOperandCastBefore(Instruction::BitCast,
6950 return BinaryOperator::create(
6951 cast<BinaryOperator>(SrcI)->getOpcode(), Op0c, Op1c);
6956 case Instruction::Shl:
6957 // Allow changing the sign of the source operand. Do not allow
6958 // changing the size of the shift, UNLESS the shift amount is a
6959 // constant. We must not change variable sized shifts to a smaller
6960 // size, because it is undefined to shift more bits out than exist
6962 if (DestBitSize == SrcBitSize ||
6963 (DestBitSize < SrcBitSize && isa<Constant>(Op1))) {
6964 Instruction::CastOps opcode = (DestBitSize == SrcBitSize ?
6965 Instruction::BitCast : Instruction::Trunc);
6966 Value *Op0c = InsertOperandCastBefore(opcode, Op0, DestTy, SrcI);
6967 Value *Op1c = InsertOperandCastBefore(opcode, Op1, DestTy, SrcI);
6968 return BinaryOperator::createShl(Op0c, Op1c);
6971 case Instruction::AShr:
6972 // If this is a signed shr, and if all bits shifted in are about to be
6973 // truncated off, turn it into an unsigned shr to allow greater
6975 if (DestBitSize < SrcBitSize &&
6976 isa<ConstantInt>(Op1)) {
6977 uint32_t ShiftAmt = cast<ConstantInt>(Op1)->getLimitedValue(SrcBitSize);
6978 if (SrcBitSize > ShiftAmt && SrcBitSize-ShiftAmt >= DestBitSize) {
6979 // Insert the new logical shift right.
6980 return BinaryOperator::createLShr(Op0, Op1);
6988 Instruction *InstCombiner::visitTrunc(TruncInst &CI) {
6989 if (Instruction *Result = commonIntCastTransforms(CI))
6992 Value *Src = CI.getOperand(0);
6993 const Type *Ty = CI.getType();
6994 uint32_t DestBitWidth = Ty->getPrimitiveSizeInBits();
6995 uint32_t SrcBitWidth = cast<IntegerType>(Src->getType())->getBitWidth();
6997 if (Instruction *SrcI = dyn_cast<Instruction>(Src)) {
6998 switch (SrcI->getOpcode()) {
7000 case Instruction::LShr:
7001 // We can shrink lshr to something smaller if we know the bits shifted in
7002 // are already zeros.
7003 if (ConstantInt *ShAmtV = dyn_cast<ConstantInt>(SrcI->getOperand(1))) {
7004 uint32_t ShAmt = ShAmtV->getLimitedValue(SrcBitWidth);
7006 // Get a mask for the bits shifting in.
7007 APInt Mask(APInt::getLowBitsSet(SrcBitWidth, ShAmt).shl(DestBitWidth));
7008 Value* SrcIOp0 = SrcI->getOperand(0);
7009 if (SrcI->hasOneUse() && MaskedValueIsZero(SrcIOp0, Mask)) {
7010 if (ShAmt >= DestBitWidth) // All zeros.
7011 return ReplaceInstUsesWith(CI, Constant::getNullValue(Ty));
7013 // Okay, we can shrink this. Truncate the input, then return a new
7015 Value *V1 = InsertCastBefore(Instruction::Trunc, SrcIOp0, Ty, CI);
7016 Value *V2 = InsertCastBefore(Instruction::Trunc, SrcI->getOperand(1),
7018 return BinaryOperator::createLShr(V1, V2);
7020 } else { // This is a variable shr.
7022 // Turn 'trunc (lshr X, Y) to bool' into '(X & (1 << Y)) != 0'. This is
7023 // more LLVM instructions, but allows '1 << Y' to be hoisted if
7024 // loop-invariant and CSE'd.
7025 if (CI.getType() == Type::Int1Ty && SrcI->hasOneUse()) {
7026 Value *One = ConstantInt::get(SrcI->getType(), 1);
7028 Value *V = InsertNewInstBefore(
7029 BinaryOperator::createShl(One, SrcI->getOperand(1),
7031 V = InsertNewInstBefore(BinaryOperator::createAnd(V,
7032 SrcI->getOperand(0),
7034 Value *Zero = Constant::getNullValue(V->getType());
7035 return new ICmpInst(ICmpInst::ICMP_NE, V, Zero);
7045 Instruction *InstCombiner::visitZExt(ZExtInst &CI) {
7046 // If one of the common conversion will work ..
7047 if (Instruction *Result = commonIntCastTransforms(CI))
7050 Value *Src = CI.getOperand(0);
7052 // If this is a cast of a cast
7053 if (CastInst *CSrc = dyn_cast<CastInst>(Src)) { // A->B->C cast
7054 // If this is a TRUNC followed by a ZEXT then we are dealing with integral
7055 // types and if the sizes are just right we can convert this into a logical
7056 // 'and' which will be much cheaper than the pair of casts.
7057 if (isa<TruncInst>(CSrc)) {
7058 // Get the sizes of the types involved
7059 Value *A = CSrc->getOperand(0);
7060 uint32_t SrcSize = A->getType()->getPrimitiveSizeInBits();
7061 uint32_t MidSize = CSrc->getType()->getPrimitiveSizeInBits();
7062 uint32_t DstSize = CI.getType()->getPrimitiveSizeInBits();
7063 // If we're actually extending zero bits and the trunc is a no-op
7064 if (MidSize < DstSize && SrcSize == DstSize) {
7065 // Replace both of the casts with an And of the type mask.
7066 APInt AndValue(APInt::getLowBitsSet(SrcSize, MidSize));
7067 Constant *AndConst = ConstantInt::get(AndValue);
7069 BinaryOperator::createAnd(CSrc->getOperand(0), AndConst);
7070 // Unfortunately, if the type changed, we need to cast it back.
7071 if (And->getType() != CI.getType()) {
7072 And->setName(CSrc->getName()+".mask");
7073 InsertNewInstBefore(And, CI);
7074 And = CastInst::createIntegerCast(And, CI.getType(), false/*ZExt*/);
7081 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Src)) {
7082 // If we are just checking for a icmp eq of a single bit and zext'ing it
7083 // to an integer, then shift the bit to the appropriate place and then
7084 // cast to integer to avoid the comparison.
7085 if (ConstantInt *Op1C = dyn_cast<ConstantInt>(ICI->getOperand(1))) {
7086 const APInt &Op1CV = Op1C->getValue();
7088 // zext (x <s 0) to i32 --> x>>u31 true if signbit set.
7089 // zext (x >s -1) to i32 --> (x>>u31)^1 true if signbit clear.
7090 if ((ICI->getPredicate() == ICmpInst::ICMP_SLT && Op1CV == 0) ||
7091 (ICI->getPredicate() == ICmpInst::ICMP_SGT &&Op1CV.isAllOnesValue())){
7092 Value *In = ICI->getOperand(0);
7093 Value *Sh = ConstantInt::get(In->getType(),
7094 In->getType()->getPrimitiveSizeInBits()-1);
7095 In = InsertNewInstBefore(BinaryOperator::createLShr(In, Sh,
7096 In->getName()+".lobit"),
7098 if (In->getType() != CI.getType())
7099 In = CastInst::createIntegerCast(In, CI.getType(),
7100 false/*ZExt*/, "tmp", &CI);
7102 if (ICI->getPredicate() == ICmpInst::ICMP_SGT) {
7103 Constant *One = ConstantInt::get(In->getType(), 1);
7104 In = InsertNewInstBefore(BinaryOperator::createXor(In, One,
7105 In->getName()+".not"),
7109 return ReplaceInstUsesWith(CI, In);
7114 // zext (X == 0) to i32 --> X^1 iff X has only the low bit set.
7115 // zext (X == 0) to i32 --> (X>>1)^1 iff X has only the 2nd bit set.
7116 // zext (X == 1) to i32 --> X iff X has only the low bit set.
7117 // zext (X == 2) to i32 --> X>>1 iff X has only the 2nd bit set.
7118 // zext (X != 0) to i32 --> X iff X has only the low bit set.
7119 // zext (X != 0) to i32 --> X>>1 iff X has only the 2nd bit set.
7120 // zext (X != 1) to i32 --> X^1 iff X has only the low bit set.
7121 // zext (X != 2) to i32 --> (X>>1)^1 iff X has only the 2nd bit set.
7122 if ((Op1CV == 0 || Op1CV.isPowerOf2()) &&
7123 // This only works for EQ and NE
7124 ICI->isEquality()) {
7125 // If Op1C some other power of two, convert:
7126 uint32_t BitWidth = Op1C->getType()->getBitWidth();
7127 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
7128 APInt TypeMask(APInt::getAllOnesValue(BitWidth));
7129 ComputeMaskedBits(ICI->getOperand(0), TypeMask, KnownZero, KnownOne);
7131 APInt KnownZeroMask(~KnownZero);
7132 if (KnownZeroMask.isPowerOf2()) { // Exactly 1 possible 1?
7133 bool isNE = ICI->getPredicate() == ICmpInst::ICMP_NE;
7134 if (Op1CV != 0 && (Op1CV != KnownZeroMask)) {
7135 // (X&4) == 2 --> false
7136 // (X&4) != 2 --> true
7137 Constant *Res = ConstantInt::get(Type::Int1Ty, isNE);
7138 Res = ConstantExpr::getZExt(Res, CI.getType());
7139 return ReplaceInstUsesWith(CI, Res);
7142 uint32_t ShiftAmt = KnownZeroMask.logBase2();
7143 Value *In = ICI->getOperand(0);
7145 // Perform a logical shr by shiftamt.
7146 // Insert the shift to put the result in the low bit.
7147 In = InsertNewInstBefore(
7148 BinaryOperator::createLShr(In,
7149 ConstantInt::get(In->getType(), ShiftAmt),
7150 In->getName()+".lobit"), CI);
7153 if ((Op1CV != 0) == isNE) { // Toggle the low bit.
7154 Constant *One = ConstantInt::get(In->getType(), 1);
7155 In = BinaryOperator::createXor(In, One, "tmp");
7156 InsertNewInstBefore(cast<Instruction>(In), CI);
7159 if (CI.getType() == In->getType())
7160 return ReplaceInstUsesWith(CI, In);
7162 return CastInst::createIntegerCast(In, CI.getType(), false/*ZExt*/);
7170 Instruction *InstCombiner::visitSExt(SExtInst &CI) {
7171 if (Instruction *I = commonIntCastTransforms(CI))
7174 Value *Src = CI.getOperand(0);
7176 // sext (x <s 0) -> ashr x, 31 -> all ones if signed
7177 // sext (x >s -1) -> ashr x, 31 -> all ones if not signed
7178 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Src)) {
7179 // If we are just checking for a icmp eq of a single bit and zext'ing it
7180 // to an integer, then shift the bit to the appropriate place and then
7181 // cast to integer to avoid the comparison.
7182 if (ConstantInt *Op1C = dyn_cast<ConstantInt>(ICI->getOperand(1))) {
7183 const APInt &Op1CV = Op1C->getValue();
7185 // sext (x <s 0) to i32 --> x>>s31 true if signbit set.
7186 // sext (x >s -1) to i32 --> (x>>s31)^-1 true if signbit clear.
7187 if ((ICI->getPredicate() == ICmpInst::ICMP_SLT && Op1CV == 0) ||
7188 (ICI->getPredicate() == ICmpInst::ICMP_SGT &&Op1CV.isAllOnesValue())){
7189 Value *In = ICI->getOperand(0);
7190 Value *Sh = ConstantInt::get(In->getType(),
7191 In->getType()->getPrimitiveSizeInBits()-1);
7192 In = InsertNewInstBefore(BinaryOperator::createAShr(In, Sh,
7193 In->getName()+".lobit"),
7195 if (In->getType() != CI.getType())
7196 In = CastInst::createIntegerCast(In, CI.getType(),
7197 true/*SExt*/, "tmp", &CI);
7199 if (ICI->getPredicate() == ICmpInst::ICMP_SGT)
7200 In = InsertNewInstBefore(BinaryOperator::createNot(In,
7201 In->getName()+".not"), CI);
7203 return ReplaceInstUsesWith(CI, In);
7211 /// FitsInFPType - Return a Constant* for the specified FP constant if it fits
7212 /// in the specified FP type without changing its value.
7213 static Constant *FitsInFPType(ConstantFP *CFP, const Type *FPTy,
7214 const fltSemantics &Sem) {
7215 APFloat F = CFP->getValueAPF();
7216 if (F.convert(Sem, APFloat::rmNearestTiesToEven) == APFloat::opOK)
7217 return ConstantFP::get(FPTy, F);
7221 /// LookThroughFPExtensions - If this is an fp extension instruction, look
7222 /// through it until we get the source value.
7223 static Value *LookThroughFPExtensions(Value *V) {
7224 if (Instruction *I = dyn_cast<Instruction>(V))
7225 if (I->getOpcode() == Instruction::FPExt)
7226 return LookThroughFPExtensions(I->getOperand(0));
7228 // If this value is a constant, return the constant in the smallest FP type
7229 // that can accurately represent it. This allows us to turn
7230 // (float)((double)X+2.0) into x+2.0f.
7231 if (ConstantFP *CFP = dyn_cast<ConstantFP>(V)) {
7232 if (CFP->getType() == Type::PPC_FP128Ty)
7233 return V; // No constant folding of this.
7234 // See if the value can be truncated to float and then reextended.
7235 if (Value *V = FitsInFPType(CFP, Type::FloatTy, APFloat::IEEEsingle))
7237 if (CFP->getType() == Type::DoubleTy)
7238 return V; // Won't shrink.
7239 if (Value *V = FitsInFPType(CFP, Type::DoubleTy, APFloat::IEEEdouble))
7241 // Don't try to shrink to various long double types.
7247 Instruction *InstCombiner::visitFPTrunc(FPTruncInst &CI) {
7248 if (Instruction *I = commonCastTransforms(CI))
7251 // If we have fptrunc(add (fpextend x), (fpextend y)), where x and y are
7252 // smaller than the destination type, we can eliminate the truncate by doing
7253 // the add as the smaller type. This applies to add/sub/mul/div as well as
7254 // many builtins (sqrt, etc).
7255 BinaryOperator *OpI = dyn_cast<BinaryOperator>(CI.getOperand(0));
7256 if (OpI && OpI->hasOneUse()) {
7257 switch (OpI->getOpcode()) {
7259 case Instruction::Add:
7260 case Instruction::Sub:
7261 case Instruction::Mul:
7262 case Instruction::FDiv:
7263 case Instruction::FRem:
7264 const Type *SrcTy = OpI->getType();
7265 Value *LHSTrunc = LookThroughFPExtensions(OpI->getOperand(0));
7266 Value *RHSTrunc = LookThroughFPExtensions(OpI->getOperand(1));
7267 if (LHSTrunc->getType() != SrcTy &&
7268 RHSTrunc->getType() != SrcTy) {
7269 unsigned DstSize = CI.getType()->getPrimitiveSizeInBits();
7270 // If the source types were both smaller than the destination type of
7271 // the cast, do this xform.
7272 if (LHSTrunc->getType()->getPrimitiveSizeInBits() <= DstSize &&
7273 RHSTrunc->getType()->getPrimitiveSizeInBits() <= DstSize) {
7274 LHSTrunc = InsertCastBefore(Instruction::FPExt, LHSTrunc,
7276 RHSTrunc = InsertCastBefore(Instruction::FPExt, RHSTrunc,
7278 return BinaryOperator::create(OpI->getOpcode(), LHSTrunc, RHSTrunc);
7287 Instruction *InstCombiner::visitFPExt(CastInst &CI) {
7288 return commonCastTransforms(CI);
7291 Instruction *InstCombiner::visitFPToUI(CastInst &CI) {
7292 return commonCastTransforms(CI);
7295 Instruction *InstCombiner::visitFPToSI(CastInst &CI) {
7296 return commonCastTransforms(CI);
7299 Instruction *InstCombiner::visitUIToFP(CastInst &CI) {
7300 return commonCastTransforms(CI);
7303 Instruction *InstCombiner::visitSIToFP(CastInst &CI) {
7304 return commonCastTransforms(CI);
7307 Instruction *InstCombiner::visitPtrToInt(CastInst &CI) {
7308 return commonPointerCastTransforms(CI);
7311 Instruction *InstCombiner::visitIntToPtr(IntToPtrInst &CI) {
7312 if (Instruction *I = commonCastTransforms(CI))
7315 const Type *DestPointee = cast<PointerType>(CI.getType())->getElementType();
7316 if (!DestPointee->isSized()) return 0;
7318 // If this is inttoptr(add (ptrtoint x), cst), try to turn this into a GEP.
7321 if (match(CI.getOperand(0), m_Add(m_Cast<PtrToIntInst>(m_Value(X)),
7322 m_ConstantInt(Cst)))) {
7323 // If the source and destination operands have the same type, see if this
7324 // is a single-index GEP.
7325 if (X->getType() == CI.getType()) {
7326 // Get the size of the pointee type.
7327 uint64_t Size = TD->getABITypeSizeInBits(DestPointee);
7329 // Convert the constant to intptr type.
7330 APInt Offset = Cst->getValue();
7331 Offset.sextOrTrunc(TD->getPointerSizeInBits());
7333 // If Offset is evenly divisible by Size, we can do this xform.
7334 if (Size && !APIntOps::srem(Offset, APInt(Offset.getBitWidth(), Size))){
7335 Offset = APIntOps::sdiv(Offset, APInt(Offset.getBitWidth(), Size));
7336 return new GetElementPtrInst(X, ConstantInt::get(Offset));
7339 // TODO: Could handle other cases, e.g. where add is indexing into field of
7341 } else if (CI.getOperand(0)->hasOneUse() &&
7342 match(CI.getOperand(0), m_Add(m_Value(X), m_ConstantInt(Cst)))) {
7343 // Otherwise, if this is inttoptr(add x, cst), try to turn this into an
7344 // "inttoptr+GEP" instead of "add+intptr".
7346 // Get the size of the pointee type.
7347 uint64_t Size = TD->getABITypeSize(DestPointee);
7349 // Convert the constant to intptr type.
7350 APInt Offset = Cst->getValue();
7351 Offset.sextOrTrunc(TD->getPointerSizeInBits());
7353 // If Offset is evenly divisible by Size, we can do this xform.
7354 if (Size && !APIntOps::srem(Offset, APInt(Offset.getBitWidth(), Size))){
7355 Offset = APIntOps::sdiv(Offset, APInt(Offset.getBitWidth(), Size));
7357 Instruction *P = InsertNewInstBefore(new IntToPtrInst(X, CI.getType(),
7359 return new GetElementPtrInst(P, ConstantInt::get(Offset), "tmp");
7365 Instruction *InstCombiner::visitBitCast(BitCastInst &CI) {
7366 // If the operands are integer typed then apply the integer transforms,
7367 // otherwise just apply the common ones.
7368 Value *Src = CI.getOperand(0);
7369 const Type *SrcTy = Src->getType();
7370 const Type *DestTy = CI.getType();
7372 if (SrcTy->isInteger() && DestTy->isInteger()) {
7373 if (Instruction *Result = commonIntCastTransforms(CI))
7375 } else if (isa<PointerType>(SrcTy)) {
7376 if (Instruction *I = commonPointerCastTransforms(CI))
7379 if (Instruction *Result = commonCastTransforms(CI))
7384 // Get rid of casts from one type to the same type. These are useless and can
7385 // be replaced by the operand.
7386 if (DestTy == Src->getType())
7387 return ReplaceInstUsesWith(CI, Src);
7389 if (const PointerType *DstPTy = dyn_cast<PointerType>(DestTy)) {
7390 const PointerType *SrcPTy = cast<PointerType>(SrcTy);
7391 const Type *DstElTy = DstPTy->getElementType();
7392 const Type *SrcElTy = SrcPTy->getElementType();
7394 // If we are casting a malloc or alloca to a pointer to a type of the same
7395 // size, rewrite the allocation instruction to allocate the "right" type.
7396 if (AllocationInst *AI = dyn_cast<AllocationInst>(Src))
7397 if (Instruction *V = PromoteCastOfAllocation(CI, *AI))
7400 // If the source and destination are pointers, and this cast is equivalent
7401 // to a getelementptr X, 0, 0, 0... turn it into the appropriate gep.
7402 // This can enhance SROA and other transforms that want type-safe pointers.
7403 Constant *ZeroUInt = Constant::getNullValue(Type::Int32Ty);
7404 unsigned NumZeros = 0;
7405 while (SrcElTy != DstElTy &&
7406 isa<CompositeType>(SrcElTy) && !isa<PointerType>(SrcElTy) &&
7407 SrcElTy->getNumContainedTypes() /* not "{}" */) {
7408 SrcElTy = cast<CompositeType>(SrcElTy)->getTypeAtIndex(ZeroUInt);
7412 // If we found a path from the src to dest, create the getelementptr now.
7413 if (SrcElTy == DstElTy) {
7414 SmallVector<Value*, 8> Idxs(NumZeros+1, ZeroUInt);
7415 return new GetElementPtrInst(Src, Idxs.begin(), Idxs.end(), "",
7416 ((Instruction*) NULL));
7420 if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(Src)) {
7421 if (SVI->hasOneUse()) {
7422 // Okay, we have (bitconvert (shuffle ..)). Check to see if this is
7423 // a bitconvert to a vector with the same # elts.
7424 if (isa<VectorType>(DestTy) &&
7425 cast<VectorType>(DestTy)->getNumElements() ==
7426 SVI->getType()->getNumElements()) {
7428 // If either of the operands is a cast from CI.getType(), then
7429 // evaluating the shuffle in the casted destination's type will allow
7430 // us to eliminate at least one cast.
7431 if (((Tmp = dyn_cast<CastInst>(SVI->getOperand(0))) &&
7432 Tmp->getOperand(0)->getType() == DestTy) ||
7433 ((Tmp = dyn_cast<CastInst>(SVI->getOperand(1))) &&
7434 Tmp->getOperand(0)->getType() == DestTy)) {
7435 Value *LHS = InsertOperandCastBefore(Instruction::BitCast,
7436 SVI->getOperand(0), DestTy, &CI);
7437 Value *RHS = InsertOperandCastBefore(Instruction::BitCast,
7438 SVI->getOperand(1), DestTy, &CI);
7439 // Return a new shuffle vector. Use the same element ID's, as we
7440 // know the vector types match #elts.
7441 return new ShuffleVectorInst(LHS, RHS, SVI->getOperand(2));
7449 /// GetSelectFoldableOperands - We want to turn code that looks like this:
7451 /// %D = select %cond, %C, %A
7453 /// %C = select %cond, %B, 0
7456 /// Assuming that the specified instruction is an operand to the select, return
7457 /// a bitmask indicating which operands of this instruction are foldable if they
7458 /// equal the other incoming value of the select.
7460 static unsigned GetSelectFoldableOperands(Instruction *I) {
7461 switch (I->getOpcode()) {
7462 case Instruction::Add:
7463 case Instruction::Mul:
7464 case Instruction::And:
7465 case Instruction::Or:
7466 case Instruction::Xor:
7467 return 3; // Can fold through either operand.
7468 case Instruction::Sub: // Can only fold on the amount subtracted.
7469 case Instruction::Shl: // Can only fold on the shift amount.
7470 case Instruction::LShr:
7471 case Instruction::AShr:
7474 return 0; // Cannot fold
7478 /// GetSelectFoldableConstant - For the same transformation as the previous
7479 /// function, return the identity constant that goes into the select.
7480 static Constant *GetSelectFoldableConstant(Instruction *I) {
7481 switch (I->getOpcode()) {
7482 default: assert(0 && "This cannot happen!"); abort();
7483 case Instruction::Add:
7484 case Instruction::Sub:
7485 case Instruction::Or:
7486 case Instruction::Xor:
7487 case Instruction::Shl:
7488 case Instruction::LShr:
7489 case Instruction::AShr:
7490 return Constant::getNullValue(I->getType());
7491 case Instruction::And:
7492 return Constant::getAllOnesValue(I->getType());
7493 case Instruction::Mul:
7494 return ConstantInt::get(I->getType(), 1);
7498 /// FoldSelectOpOp - Here we have (select c, TI, FI), and we know that TI and FI
7499 /// have the same opcode and only one use each. Try to simplify this.
7500 Instruction *InstCombiner::FoldSelectOpOp(SelectInst &SI, Instruction *TI,
7502 if (TI->getNumOperands() == 1) {
7503 // If this is a non-volatile load or a cast from the same type,
7506 if (TI->getOperand(0)->getType() != FI->getOperand(0)->getType())
7509 return 0; // unknown unary op.
7512 // Fold this by inserting a select from the input values.
7513 SelectInst *NewSI = new SelectInst(SI.getCondition(), TI->getOperand(0),
7514 FI->getOperand(0), SI.getName()+".v");
7515 InsertNewInstBefore(NewSI, SI);
7516 return CastInst::create(Instruction::CastOps(TI->getOpcode()), NewSI,
7520 // Only handle binary operators here.
7521 if (!isa<BinaryOperator>(TI))
7524 // Figure out if the operations have any operands in common.
7525 Value *MatchOp, *OtherOpT, *OtherOpF;
7527 if (TI->getOperand(0) == FI->getOperand(0)) {
7528 MatchOp = TI->getOperand(0);
7529 OtherOpT = TI->getOperand(1);
7530 OtherOpF = FI->getOperand(1);
7531 MatchIsOpZero = true;
7532 } else if (TI->getOperand(1) == FI->getOperand(1)) {
7533 MatchOp = TI->getOperand(1);
7534 OtherOpT = TI->getOperand(0);
7535 OtherOpF = FI->getOperand(0);
7536 MatchIsOpZero = false;
7537 } else if (!TI->isCommutative()) {
7539 } else if (TI->getOperand(0) == FI->getOperand(1)) {
7540 MatchOp = TI->getOperand(0);
7541 OtherOpT = TI->getOperand(1);
7542 OtherOpF = FI->getOperand(0);
7543 MatchIsOpZero = true;
7544 } else if (TI->getOperand(1) == FI->getOperand(0)) {
7545 MatchOp = TI->getOperand(1);
7546 OtherOpT = TI->getOperand(0);
7547 OtherOpF = FI->getOperand(1);
7548 MatchIsOpZero = true;
7553 // If we reach here, they do have operations in common.
7554 SelectInst *NewSI = new SelectInst(SI.getCondition(), OtherOpT,
7555 OtherOpF, SI.getName()+".v");
7556 InsertNewInstBefore(NewSI, SI);
7558 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(TI)) {
7560 return BinaryOperator::create(BO->getOpcode(), MatchOp, NewSI);
7562 return BinaryOperator::create(BO->getOpcode(), NewSI, MatchOp);
7564 assert(0 && "Shouldn't get here");
7568 Instruction *InstCombiner::visitSelectInst(SelectInst &SI) {
7569 Value *CondVal = SI.getCondition();
7570 Value *TrueVal = SI.getTrueValue();
7571 Value *FalseVal = SI.getFalseValue();
7573 // select true, X, Y -> X
7574 // select false, X, Y -> Y
7575 if (ConstantInt *C = dyn_cast<ConstantInt>(CondVal))
7576 return ReplaceInstUsesWith(SI, C->getZExtValue() ? TrueVal : FalseVal);
7578 // select C, X, X -> X
7579 if (TrueVal == FalseVal)
7580 return ReplaceInstUsesWith(SI, TrueVal);
7582 if (isa<UndefValue>(TrueVal)) // select C, undef, X -> X
7583 return ReplaceInstUsesWith(SI, FalseVal);
7584 if (isa<UndefValue>(FalseVal)) // select C, X, undef -> X
7585 return ReplaceInstUsesWith(SI, TrueVal);
7586 if (isa<UndefValue>(CondVal)) { // select undef, X, Y -> X or Y
7587 if (isa<Constant>(TrueVal))
7588 return ReplaceInstUsesWith(SI, TrueVal);
7590 return ReplaceInstUsesWith(SI, FalseVal);
7593 if (SI.getType() == Type::Int1Ty) {
7594 if (ConstantInt *C = dyn_cast<ConstantInt>(TrueVal)) {
7595 if (C->getZExtValue()) {
7596 // Change: A = select B, true, C --> A = or B, C
7597 return BinaryOperator::createOr(CondVal, FalseVal);
7599 // Change: A = select B, false, C --> A = and !B, C
7601 InsertNewInstBefore(BinaryOperator::createNot(CondVal,
7602 "not."+CondVal->getName()), SI);
7603 return BinaryOperator::createAnd(NotCond, FalseVal);
7605 } else if (ConstantInt *C = dyn_cast<ConstantInt>(FalseVal)) {
7606 if (C->getZExtValue() == false) {
7607 // Change: A = select B, C, false --> A = and B, C
7608 return BinaryOperator::createAnd(CondVal, TrueVal);
7610 // Change: A = select B, C, true --> A = or !B, C
7612 InsertNewInstBefore(BinaryOperator::createNot(CondVal,
7613 "not."+CondVal->getName()), SI);
7614 return BinaryOperator::createOr(NotCond, TrueVal);
7618 // select a, b, a -> a&b
7619 // select a, a, b -> a|b
7620 if (CondVal == TrueVal)
7621 return BinaryOperator::createOr(CondVal, FalseVal);
7622 else if (CondVal == FalseVal)
7623 return BinaryOperator::createAnd(CondVal, TrueVal);
7626 // Selecting between two integer constants?
7627 if (ConstantInt *TrueValC = dyn_cast<ConstantInt>(TrueVal))
7628 if (ConstantInt *FalseValC = dyn_cast<ConstantInt>(FalseVal)) {
7629 // select C, 1, 0 -> zext C to int
7630 if (FalseValC->isZero() && TrueValC->getValue() == 1) {
7631 return CastInst::create(Instruction::ZExt, CondVal, SI.getType());
7632 } else if (TrueValC->isZero() && FalseValC->getValue() == 1) {
7633 // select C, 0, 1 -> zext !C to int
7635 InsertNewInstBefore(BinaryOperator::createNot(CondVal,
7636 "not."+CondVal->getName()), SI);
7637 return CastInst::create(Instruction::ZExt, NotCond, SI.getType());
7640 // FIXME: Turn select 0/-1 and -1/0 into sext from condition!
7642 if (ICmpInst *IC = dyn_cast<ICmpInst>(SI.getCondition())) {
7644 // (x <s 0) ? -1 : 0 -> ashr x, 31
7645 if (TrueValC->isAllOnesValue() && FalseValC->isZero())
7646 if (ConstantInt *CmpCst = dyn_cast<ConstantInt>(IC->getOperand(1))) {
7647 if (IC->getPredicate() == ICmpInst::ICMP_SLT && CmpCst->isZero()) {
7648 // The comparison constant and the result are not neccessarily the
7649 // same width. Make an all-ones value by inserting a AShr.
7650 Value *X = IC->getOperand(0);
7651 uint32_t Bits = X->getType()->getPrimitiveSizeInBits();
7652 Constant *ShAmt = ConstantInt::get(X->getType(), Bits-1);
7653 Instruction *SRA = BinaryOperator::create(Instruction::AShr, X,
7655 InsertNewInstBefore(SRA, SI);
7657 // Finally, convert to the type of the select RHS. We figure out
7658 // if this requires a SExt, Trunc or BitCast based on the sizes.
7659 Instruction::CastOps opc = Instruction::BitCast;
7660 uint32_t SRASize = SRA->getType()->getPrimitiveSizeInBits();
7661 uint32_t SISize = SI.getType()->getPrimitiveSizeInBits();
7662 if (SRASize < SISize)
7663 opc = Instruction::SExt;
7664 else if (SRASize > SISize)
7665 opc = Instruction::Trunc;
7666 return CastInst::create(opc, SRA, SI.getType());
7671 // If one of the constants is zero (we know they can't both be) and we
7672 // have an icmp instruction with zero, and we have an 'and' with the
7673 // non-constant value, eliminate this whole mess. This corresponds to
7674 // cases like this: ((X & 27) ? 27 : 0)
7675 if (TrueValC->isZero() || FalseValC->isZero())
7676 if (IC->isEquality() && isa<ConstantInt>(IC->getOperand(1)) &&
7677 cast<Constant>(IC->getOperand(1))->isNullValue())
7678 if (Instruction *ICA = dyn_cast<Instruction>(IC->getOperand(0)))
7679 if (ICA->getOpcode() == Instruction::And &&
7680 isa<ConstantInt>(ICA->getOperand(1)) &&
7681 (ICA->getOperand(1) == TrueValC ||
7682 ICA->getOperand(1) == FalseValC) &&
7683 isOneBitSet(cast<ConstantInt>(ICA->getOperand(1)))) {
7684 // Okay, now we know that everything is set up, we just don't
7685 // know whether we have a icmp_ne or icmp_eq and whether the
7686 // true or false val is the zero.
7687 bool ShouldNotVal = !TrueValC->isZero();
7688 ShouldNotVal ^= IC->getPredicate() == ICmpInst::ICMP_NE;
7691 V = InsertNewInstBefore(BinaryOperator::create(
7692 Instruction::Xor, V, ICA->getOperand(1)), SI);
7693 return ReplaceInstUsesWith(SI, V);
7698 // See if we are selecting two values based on a comparison of the two values.
7699 if (FCmpInst *FCI = dyn_cast<FCmpInst>(CondVal)) {
7700 if (FCI->getOperand(0) == TrueVal && FCI->getOperand(1) == FalseVal) {
7701 // Transform (X == Y) ? X : Y -> Y
7702 if (FCI->getPredicate() == FCmpInst::FCMP_OEQ) {
7703 // This is not safe in general for floating point:
7704 // consider X== -0, Y== +0.
7705 // It becomes safe if either operand is a nonzero constant.
7706 ConstantFP *CFPt, *CFPf;
7707 if (((CFPt = dyn_cast<ConstantFP>(TrueVal)) &&
7708 !CFPt->getValueAPF().isZero()) ||
7709 ((CFPf = dyn_cast<ConstantFP>(FalseVal)) &&
7710 !CFPf->getValueAPF().isZero()))
7711 return ReplaceInstUsesWith(SI, FalseVal);
7713 // Transform (X != Y) ? X : Y -> X
7714 if (FCI->getPredicate() == FCmpInst::FCMP_ONE)
7715 return ReplaceInstUsesWith(SI, TrueVal);
7716 // NOTE: if we wanted to, this is where to detect MIN/MAX/ABS/etc.
7718 } else if (FCI->getOperand(0) == FalseVal && FCI->getOperand(1) == TrueVal){
7719 // Transform (X == Y) ? Y : X -> X
7720 if (FCI->getPredicate() == FCmpInst::FCMP_OEQ) {
7721 // This is not safe in general for floating point:
7722 // consider X== -0, Y== +0.
7723 // It becomes safe if either operand is a nonzero constant.
7724 ConstantFP *CFPt, *CFPf;
7725 if (((CFPt = dyn_cast<ConstantFP>(TrueVal)) &&
7726 !CFPt->getValueAPF().isZero()) ||
7727 ((CFPf = dyn_cast<ConstantFP>(FalseVal)) &&
7728 !CFPf->getValueAPF().isZero()))
7729 return ReplaceInstUsesWith(SI, FalseVal);
7731 // Transform (X != Y) ? Y : X -> Y
7732 if (FCI->getPredicate() == FCmpInst::FCMP_ONE)
7733 return ReplaceInstUsesWith(SI, TrueVal);
7734 // NOTE: if we wanted to, this is where to detect MIN/MAX/ABS/etc.
7738 // See if we are selecting two values based on a comparison of the two values.
7739 if (ICmpInst *ICI = dyn_cast<ICmpInst>(CondVal)) {
7740 if (ICI->getOperand(0) == TrueVal && ICI->getOperand(1) == FalseVal) {
7741 // Transform (X == Y) ? X : Y -> Y
7742 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
7743 return ReplaceInstUsesWith(SI, FalseVal);
7744 // Transform (X != Y) ? X : Y -> X
7745 if (ICI->getPredicate() == ICmpInst::ICMP_NE)
7746 return ReplaceInstUsesWith(SI, TrueVal);
7747 // NOTE: if we wanted to, this is where to detect MIN/MAX/ABS/etc.
7749 } else if (ICI->getOperand(0) == FalseVal && ICI->getOperand(1) == TrueVal){
7750 // Transform (X == Y) ? Y : X -> X
7751 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
7752 return ReplaceInstUsesWith(SI, FalseVal);
7753 // Transform (X != Y) ? Y : X -> Y
7754 if (ICI->getPredicate() == ICmpInst::ICMP_NE)
7755 return ReplaceInstUsesWith(SI, TrueVal);
7756 // NOTE: if we wanted to, this is where to detect MIN/MAX/ABS/etc.
7760 if (Instruction *TI = dyn_cast<Instruction>(TrueVal))
7761 if (Instruction *FI = dyn_cast<Instruction>(FalseVal))
7762 if (TI->hasOneUse() && FI->hasOneUse()) {
7763 Instruction *AddOp = 0, *SubOp = 0;
7765 // Turn (select C, (op X, Y), (op X, Z)) -> (op X, (select C, Y, Z))
7766 if (TI->getOpcode() == FI->getOpcode())
7767 if (Instruction *IV = FoldSelectOpOp(SI, TI, FI))
7770 // Turn select C, (X+Y), (X-Y) --> (X+(select C, Y, (-Y))). This is
7771 // even legal for FP.
7772 if (TI->getOpcode() == Instruction::Sub &&
7773 FI->getOpcode() == Instruction::Add) {
7774 AddOp = FI; SubOp = TI;
7775 } else if (FI->getOpcode() == Instruction::Sub &&
7776 TI->getOpcode() == Instruction::Add) {
7777 AddOp = TI; SubOp = FI;
7781 Value *OtherAddOp = 0;
7782 if (SubOp->getOperand(0) == AddOp->getOperand(0)) {
7783 OtherAddOp = AddOp->getOperand(1);
7784 } else if (SubOp->getOperand(0) == AddOp->getOperand(1)) {
7785 OtherAddOp = AddOp->getOperand(0);
7789 // So at this point we know we have (Y -> OtherAddOp):
7790 // select C, (add X, Y), (sub X, Z)
7791 Value *NegVal; // Compute -Z
7792 if (Constant *C = dyn_cast<Constant>(SubOp->getOperand(1))) {
7793 NegVal = ConstantExpr::getNeg(C);
7795 NegVal = InsertNewInstBefore(
7796 BinaryOperator::createNeg(SubOp->getOperand(1), "tmp"), SI);
7799 Value *NewTrueOp = OtherAddOp;
7800 Value *NewFalseOp = NegVal;
7802 std::swap(NewTrueOp, NewFalseOp);
7803 Instruction *NewSel =
7804 new SelectInst(CondVal, NewTrueOp,NewFalseOp,SI.getName()+".p");
7806 NewSel = InsertNewInstBefore(NewSel, SI);
7807 return BinaryOperator::createAdd(SubOp->getOperand(0), NewSel);
7812 // See if we can fold the select into one of our operands.
7813 if (SI.getType()->isInteger()) {
7814 // See the comment above GetSelectFoldableOperands for a description of the
7815 // transformation we are doing here.
7816 if (Instruction *TVI = dyn_cast<Instruction>(TrueVal))
7817 if (TVI->hasOneUse() && TVI->getNumOperands() == 2 &&
7818 !isa<Constant>(FalseVal))
7819 if (unsigned SFO = GetSelectFoldableOperands(TVI)) {
7820 unsigned OpToFold = 0;
7821 if ((SFO & 1) && FalseVal == TVI->getOperand(0)) {
7823 } else if ((SFO & 2) && FalseVal == TVI->getOperand(1)) {
7828 Constant *C = GetSelectFoldableConstant(TVI);
7829 Instruction *NewSel =
7830 new SelectInst(SI.getCondition(), TVI->getOperand(2-OpToFold), C);
7831 InsertNewInstBefore(NewSel, SI);
7832 NewSel->takeName(TVI);
7833 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(TVI))
7834 return BinaryOperator::create(BO->getOpcode(), FalseVal, NewSel);
7836 assert(0 && "Unknown instruction!!");
7841 if (Instruction *FVI = dyn_cast<Instruction>(FalseVal))
7842 if (FVI->hasOneUse() && FVI->getNumOperands() == 2 &&
7843 !isa<Constant>(TrueVal))
7844 if (unsigned SFO = GetSelectFoldableOperands(FVI)) {
7845 unsigned OpToFold = 0;
7846 if ((SFO & 1) && TrueVal == FVI->getOperand(0)) {
7848 } else if ((SFO & 2) && TrueVal == FVI->getOperand(1)) {
7853 Constant *C = GetSelectFoldableConstant(FVI);
7854 Instruction *NewSel =
7855 new SelectInst(SI.getCondition(), C, FVI->getOperand(2-OpToFold));
7856 InsertNewInstBefore(NewSel, SI);
7857 NewSel->takeName(FVI);
7858 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(FVI))
7859 return BinaryOperator::create(BO->getOpcode(), TrueVal, NewSel);
7861 assert(0 && "Unknown instruction!!");
7866 if (BinaryOperator::isNot(CondVal)) {
7867 SI.setOperand(0, BinaryOperator::getNotArgument(CondVal));
7868 SI.setOperand(1, FalseVal);
7869 SI.setOperand(2, TrueVal);
7876 /// GetOrEnforceKnownAlignment - If the specified pointer has an alignment that
7877 /// we can determine, return it, otherwise return 0. If PrefAlign is specified,
7878 /// and it is more than the alignment of the ultimate object, see if we can
7879 /// increase the alignment of the ultimate object, making this check succeed.
7880 static unsigned GetOrEnforceKnownAlignment(Value *V, TargetData *TD,
7881 unsigned PrefAlign = 0) {
7882 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(V)) {
7883 unsigned Align = GV->getAlignment();
7884 if (Align == 0 && TD && GV->getType()->getElementType()->isSized())
7885 Align = TD->getPrefTypeAlignment(GV->getType()->getElementType());
7887 // If there is a large requested alignment and we can, bump up the alignment
7889 if (PrefAlign > Align && GV->hasInitializer()) {
7890 GV->setAlignment(PrefAlign);
7894 } else if (AllocationInst *AI = dyn_cast<AllocationInst>(V)) {
7895 unsigned Align = AI->getAlignment();
7896 if (Align == 0 && TD) {
7897 if (isa<AllocaInst>(AI))
7898 Align = TD->getPrefTypeAlignment(AI->getType()->getElementType());
7899 else if (isa<MallocInst>(AI)) {
7900 // Malloc returns maximally aligned memory.
7901 Align = TD->getABITypeAlignment(AI->getType()->getElementType());
7904 (unsigned)TD->getABITypeAlignment(Type::DoubleTy));
7907 (unsigned)TD->getABITypeAlignment(Type::Int64Ty));
7911 // If there is a requested alignment and if this is an alloca, round up. We
7912 // don't do this for malloc, because some systems can't respect the request.
7913 if (PrefAlign > Align && isa<AllocaInst>(AI)) {
7914 AI->setAlignment(PrefAlign);
7918 } else if (isa<BitCastInst>(V) ||
7919 (isa<ConstantExpr>(V) &&
7920 cast<ConstantExpr>(V)->getOpcode() == Instruction::BitCast)) {
7921 return GetOrEnforceKnownAlignment(cast<User>(V)->getOperand(0),
7923 } else if (User *GEPI = dyn_castGetElementPtr(V)) {
7924 // If all indexes are zero, it is just the alignment of the base pointer.
7925 bool AllZeroOperands = true;
7926 for (unsigned i = 1, e = GEPI->getNumOperands(); i != e; ++i)
7927 if (!isa<Constant>(GEPI->getOperand(i)) ||
7928 !cast<Constant>(GEPI->getOperand(i))->isNullValue()) {
7929 AllZeroOperands = false;
7933 if (AllZeroOperands) {
7934 // Treat this like a bitcast.
7935 return GetOrEnforceKnownAlignment(GEPI->getOperand(0), TD, PrefAlign);
7938 unsigned BaseAlignment = GetOrEnforceKnownAlignment(GEPI->getOperand(0),TD);
7939 if (BaseAlignment == 0) return 0;
7941 // Otherwise, if the base alignment is >= the alignment we expect for the
7942 // base pointer type, then we know that the resultant pointer is aligned at
7943 // least as much as its type requires.
7946 const Type *BasePtrTy = GEPI->getOperand(0)->getType();
7947 const PointerType *PtrTy = cast<PointerType>(BasePtrTy);
7948 unsigned Align = TD->getABITypeAlignment(PtrTy->getElementType());
7949 if (Align <= BaseAlignment) {
7950 const Type *GEPTy = GEPI->getType();
7951 const PointerType *GEPPtrTy = cast<PointerType>(GEPTy);
7952 Align = std::min(Align, (unsigned)
7953 TD->getABITypeAlignment(GEPPtrTy->getElementType()));
7961 Instruction *InstCombiner::SimplifyMemTransfer(MemIntrinsic *MI) {
7962 unsigned DstAlign = GetOrEnforceKnownAlignment(MI->getOperand(1), TD);
7963 unsigned SrcAlign = GetOrEnforceKnownAlignment(MI->getOperand(2), TD);
7964 unsigned MinAlign = std::min(DstAlign, SrcAlign);
7965 unsigned CopyAlign = MI->getAlignment()->getZExtValue();
7967 if (CopyAlign < MinAlign) {
7968 MI->setAlignment(ConstantInt::get(Type::Int32Ty, MinAlign));
7972 // If MemCpyInst length is 1/2/4/8 bytes then replace memcpy with
7974 ConstantInt *MemOpLength = dyn_cast<ConstantInt>(MI->getOperand(3));
7975 if (MemOpLength == 0) return 0;
7977 // Source and destination pointer types are always "i8*" for intrinsic. See
7978 // if the size is something we can handle with a single primitive load/store.
7979 // A single load+store correctly handles overlapping memory in the memmove
7981 unsigned Size = MemOpLength->getZExtValue();
7982 if (Size == 0 || Size > 8 || (Size&(Size-1)))
7983 return 0; // If not 1/2/4/8 bytes, exit.
7985 // Use an integer load+store unless we can find something better.
7986 Type *NewPtrTy = PointerType::getUnqual(IntegerType::get(Size<<3));
7988 // Memcpy forces the use of i8* for the source and destination. That means
7989 // that if you're using memcpy to move one double around, you'll get a cast
7990 // from double* to i8*. We'd much rather use a double load+store rather than
7991 // an i64 load+store, here because this improves the odds that the source or
7992 // dest address will be promotable. See if we can find a better type than the
7993 // integer datatype.
7994 if (Value *Op = getBitCastOperand(MI->getOperand(1))) {
7995 const Type *SrcETy = cast<PointerType>(Op->getType())->getElementType();
7996 if (SrcETy->isSized() && TD->getTypeStoreSize(SrcETy) == Size) {
7997 // The SrcETy might be something like {{{double}}} or [1 x double]. Rip
7998 // down through these levels if so.
7999 while (!SrcETy->isFirstClassType()) {
8000 if (const StructType *STy = dyn_cast<StructType>(SrcETy)) {
8001 if (STy->getNumElements() == 1)
8002 SrcETy = STy->getElementType(0);
8005 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(SrcETy)) {
8006 if (ATy->getNumElements() == 1)
8007 SrcETy = ATy->getElementType();
8014 if (SrcETy->isFirstClassType())
8015 NewPtrTy = PointerType::getUnqual(SrcETy);
8020 // If the memcpy/memmove provides better alignment info than we can
8022 SrcAlign = std::max(SrcAlign, CopyAlign);
8023 DstAlign = std::max(DstAlign, CopyAlign);
8025 Value *Src = InsertBitCastBefore(MI->getOperand(2), NewPtrTy, *MI);
8026 Value *Dest = InsertBitCastBefore(MI->getOperand(1), NewPtrTy, *MI);
8027 Instruction *L = new LoadInst(Src, "tmp", false, SrcAlign);
8028 InsertNewInstBefore(L, *MI);
8029 InsertNewInstBefore(new StoreInst(L, Dest, false, DstAlign), *MI);
8031 // Set the size of the copy to 0, it will be deleted on the next iteration.
8032 MI->setOperand(3, Constant::getNullValue(MemOpLength->getType()));
8036 /// visitCallInst - CallInst simplification. This mostly only handles folding
8037 /// of intrinsic instructions. For normal calls, it allows visitCallSite to do
8038 /// the heavy lifting.
8040 Instruction *InstCombiner::visitCallInst(CallInst &CI) {
8041 IntrinsicInst *II = dyn_cast<IntrinsicInst>(&CI);
8042 if (!II) return visitCallSite(&CI);
8044 // Intrinsics cannot occur in an invoke, so handle them here instead of in
8046 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(II)) {
8047 bool Changed = false;
8049 // memmove/cpy/set of zero bytes is a noop.
8050 if (Constant *NumBytes = dyn_cast<Constant>(MI->getLength())) {
8051 if (NumBytes->isNullValue()) return EraseInstFromFunction(CI);
8053 if (ConstantInt *CI = dyn_cast<ConstantInt>(NumBytes))
8054 if (CI->getZExtValue() == 1) {
8055 // Replace the instruction with just byte operations. We would
8056 // transform other cases to loads/stores, but we don't know if
8057 // alignment is sufficient.
8061 // If we have a memmove and the source operation is a constant global,
8062 // then the source and dest pointers can't alias, so we can change this
8063 // into a call to memcpy.
8064 if (MemMoveInst *MMI = dyn_cast<MemMoveInst>(MI)) {
8065 if (GlobalVariable *GVSrc = dyn_cast<GlobalVariable>(MMI->getSource()))
8066 if (GVSrc->isConstant()) {
8067 Module *M = CI.getParent()->getParent()->getParent();
8068 Intrinsic::ID MemCpyID;
8069 if (CI.getOperand(3)->getType() == Type::Int32Ty)
8070 MemCpyID = Intrinsic::memcpy_i32;
8072 MemCpyID = Intrinsic::memcpy_i64;
8073 CI.setOperand(0, Intrinsic::getDeclaration(M, MemCpyID));
8078 // If we can determine a pointer alignment that is bigger than currently
8079 // set, update the alignment.
8080 if (isa<MemCpyInst>(MI) || isa<MemMoveInst>(MI)) {
8081 if (Instruction *I = SimplifyMemTransfer(MI))
8083 } else if (isa<MemSetInst>(MI)) {
8084 unsigned Alignment = GetOrEnforceKnownAlignment(MI->getDest(), TD);
8085 if (MI->getAlignment()->getZExtValue() < Alignment) {
8086 MI->setAlignment(ConstantInt::get(Type::Int32Ty, Alignment));
8091 if (Changed) return II;
8093 switch (II->getIntrinsicID()) {
8095 case Intrinsic::ppc_altivec_lvx:
8096 case Intrinsic::ppc_altivec_lvxl:
8097 case Intrinsic::x86_sse_loadu_ps:
8098 case Intrinsic::x86_sse2_loadu_pd:
8099 case Intrinsic::x86_sse2_loadu_dq:
8100 // Turn PPC lvx -> load if the pointer is known aligned.
8101 // Turn X86 loadups -> load if the pointer is known aligned.
8102 if (GetOrEnforceKnownAlignment(II->getOperand(1), TD, 16) >= 16) {
8103 Value *Ptr = InsertBitCastBefore(II->getOperand(1),
8104 PointerType::getUnqual(II->getType()),
8106 return new LoadInst(Ptr);
8109 case Intrinsic::ppc_altivec_stvx:
8110 case Intrinsic::ppc_altivec_stvxl:
8111 // Turn stvx -> store if the pointer is known aligned.
8112 if (GetOrEnforceKnownAlignment(II->getOperand(2), TD, 16) >= 16) {
8113 const Type *OpPtrTy =
8114 PointerType::getUnqual(II->getOperand(1)->getType());
8115 Value *Ptr = InsertBitCastBefore(II->getOperand(2), OpPtrTy, CI);
8116 return new StoreInst(II->getOperand(1), Ptr);
8119 case Intrinsic::x86_sse_storeu_ps:
8120 case Intrinsic::x86_sse2_storeu_pd:
8121 case Intrinsic::x86_sse2_storeu_dq:
8122 case Intrinsic::x86_sse2_storel_dq:
8123 // Turn X86 storeu -> store if the pointer is known aligned.
8124 if (GetOrEnforceKnownAlignment(II->getOperand(1), TD, 16) >= 16) {
8125 const Type *OpPtrTy =
8126 PointerType::getUnqual(II->getOperand(2)->getType());
8127 Value *Ptr = InsertBitCastBefore(II->getOperand(1), OpPtrTy, CI);
8128 return new StoreInst(II->getOperand(2), Ptr);
8132 case Intrinsic::x86_sse_cvttss2si: {
8133 // These intrinsics only demands the 0th element of its input vector. If
8134 // we can simplify the input based on that, do so now.
8136 if (Value *V = SimplifyDemandedVectorElts(II->getOperand(1), 1,
8138 II->setOperand(1, V);
8144 case Intrinsic::ppc_altivec_vperm:
8145 // Turn vperm(V1,V2,mask) -> shuffle(V1,V2,mask) if mask is a constant.
8146 if (ConstantVector *Mask = dyn_cast<ConstantVector>(II->getOperand(3))) {
8147 assert(Mask->getNumOperands() == 16 && "Bad type for intrinsic!");
8149 // Check that all of the elements are integer constants or undefs.
8150 bool AllEltsOk = true;
8151 for (unsigned i = 0; i != 16; ++i) {
8152 if (!isa<ConstantInt>(Mask->getOperand(i)) &&
8153 !isa<UndefValue>(Mask->getOperand(i))) {
8160 // Cast the input vectors to byte vectors.
8161 Value *Op0 =InsertBitCastBefore(II->getOperand(1),Mask->getType(),CI);
8162 Value *Op1 =InsertBitCastBefore(II->getOperand(2),Mask->getType(),CI);
8163 Value *Result = UndefValue::get(Op0->getType());
8165 // Only extract each element once.
8166 Value *ExtractedElts[32];
8167 memset(ExtractedElts, 0, sizeof(ExtractedElts));
8169 for (unsigned i = 0; i != 16; ++i) {
8170 if (isa<UndefValue>(Mask->getOperand(i)))
8172 unsigned Idx=cast<ConstantInt>(Mask->getOperand(i))->getZExtValue();
8173 Idx &= 31; // Match the hardware behavior.
8175 if (ExtractedElts[Idx] == 0) {
8177 new ExtractElementInst(Idx < 16 ? Op0 : Op1, Idx&15, "tmp");
8178 InsertNewInstBefore(Elt, CI);
8179 ExtractedElts[Idx] = Elt;
8182 // Insert this value into the result vector.
8183 Result = new InsertElementInst(Result, ExtractedElts[Idx], i,"tmp");
8184 InsertNewInstBefore(cast<Instruction>(Result), CI);
8186 return CastInst::create(Instruction::BitCast, Result, CI.getType());
8191 case Intrinsic::stackrestore: {
8192 // If the save is right next to the restore, remove the restore. This can
8193 // happen when variable allocas are DCE'd.
8194 if (IntrinsicInst *SS = dyn_cast<IntrinsicInst>(II->getOperand(1))) {
8195 if (SS->getIntrinsicID() == Intrinsic::stacksave) {
8196 BasicBlock::iterator BI = SS;
8198 return EraseInstFromFunction(CI);
8202 // Scan down this block to see if there is another stack restore in the
8203 // same block without an intervening call/alloca.
8204 BasicBlock::iterator BI = II;
8205 TerminatorInst *TI = II->getParent()->getTerminator();
8206 bool CannotRemove = false;
8207 for (++BI; &*BI != TI; ++BI) {
8208 if (isa<AllocaInst>(BI)) {
8209 CannotRemove = true;
8212 if (isa<CallInst>(BI)) {
8213 if (!isa<IntrinsicInst>(BI)) {
8214 CannotRemove = true;
8217 // If there is a stackrestore below this one, remove this one.
8218 return EraseInstFromFunction(CI);
8222 // If the stack restore is in a return/unwind block and if there are no
8223 // allocas or calls between the restore and the return, nuke the restore.
8224 if (!CannotRemove && (isa<ReturnInst>(TI) || isa<UnwindInst>(TI)))
8225 return EraseInstFromFunction(CI);
8231 return visitCallSite(II);
8234 // InvokeInst simplification
8236 Instruction *InstCombiner::visitInvokeInst(InvokeInst &II) {
8237 return visitCallSite(&II);
8240 // visitCallSite - Improvements for call and invoke instructions.
8242 Instruction *InstCombiner::visitCallSite(CallSite CS) {
8243 bool Changed = false;
8245 // If the callee is a constexpr cast of a function, attempt to move the cast
8246 // to the arguments of the call/invoke.
8247 if (transformConstExprCastCall(CS)) return 0;
8249 Value *Callee = CS.getCalledValue();
8251 if (Function *CalleeF = dyn_cast<Function>(Callee))
8252 if (CalleeF->getCallingConv() != CS.getCallingConv()) {
8253 Instruction *OldCall = CS.getInstruction();
8254 // If the call and callee calling conventions don't match, this call must
8255 // be unreachable, as the call is undefined.
8256 new StoreInst(ConstantInt::getTrue(),
8257 UndefValue::get(PointerType::getUnqual(Type::Int1Ty)),
8259 if (!OldCall->use_empty())
8260 OldCall->replaceAllUsesWith(UndefValue::get(OldCall->getType()));
8261 if (isa<CallInst>(OldCall)) // Not worth removing an invoke here.
8262 return EraseInstFromFunction(*OldCall);
8266 if (isa<ConstantPointerNull>(Callee) || isa<UndefValue>(Callee)) {
8267 // This instruction is not reachable, just remove it. We insert a store to
8268 // undef so that we know that this code is not reachable, despite the fact
8269 // that we can't modify the CFG here.
8270 new StoreInst(ConstantInt::getTrue(),
8271 UndefValue::get(PointerType::getUnqual(Type::Int1Ty)),
8272 CS.getInstruction());
8274 if (!CS.getInstruction()->use_empty())
8275 CS.getInstruction()->
8276 replaceAllUsesWith(UndefValue::get(CS.getInstruction()->getType()));
8278 if (InvokeInst *II = dyn_cast<InvokeInst>(CS.getInstruction())) {
8279 // Don't break the CFG, insert a dummy cond branch.
8280 new BranchInst(II->getNormalDest(), II->getUnwindDest(),
8281 ConstantInt::getTrue(), II);
8283 return EraseInstFromFunction(*CS.getInstruction());
8286 if (BitCastInst *BC = dyn_cast<BitCastInst>(Callee))
8287 if (IntrinsicInst *In = dyn_cast<IntrinsicInst>(BC->getOperand(0)))
8288 if (In->getIntrinsicID() == Intrinsic::init_trampoline)
8289 return transformCallThroughTrampoline(CS);
8291 const PointerType *PTy = cast<PointerType>(Callee->getType());
8292 const FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
8293 if (FTy->isVarArg()) {
8294 // See if we can optimize any arguments passed through the varargs area of
8296 for (CallSite::arg_iterator I = CS.arg_begin()+FTy->getNumParams(),
8297 E = CS.arg_end(); I != E; ++I)
8298 if (CastInst *CI = dyn_cast<CastInst>(*I)) {
8299 // If this cast does not effect the value passed through the varargs
8300 // area, we can eliminate the use of the cast.
8301 Value *Op = CI->getOperand(0);
8302 if (CI->isLosslessCast()) {
8309 if (isa<InlineAsm>(Callee) && !CS.doesNotThrow()) {
8310 // Inline asm calls cannot throw - mark them 'nounwind'.
8311 CS.setDoesNotThrow();
8315 return Changed ? CS.getInstruction() : 0;
8318 // transformConstExprCastCall - If the callee is a constexpr cast of a function,
8319 // attempt to move the cast to the arguments of the call/invoke.
8321 bool InstCombiner::transformConstExprCastCall(CallSite CS) {
8322 if (!isa<ConstantExpr>(CS.getCalledValue())) return false;
8323 ConstantExpr *CE = cast<ConstantExpr>(CS.getCalledValue());
8324 if (CE->getOpcode() != Instruction::BitCast ||
8325 !isa<Function>(CE->getOperand(0)))
8327 Function *Callee = cast<Function>(CE->getOperand(0));
8328 Instruction *Caller = CS.getInstruction();
8329 const ParamAttrsList* CallerPAL = CS.getParamAttrs();
8331 // Okay, this is a cast from a function to a different type. Unless doing so
8332 // would cause a type conversion of one of our arguments, change this call to
8333 // be a direct call with arguments casted to the appropriate types.
8335 const FunctionType *FT = Callee->getFunctionType();
8336 const Type *OldRetTy = Caller->getType();
8338 // Check to see if we are changing the return type...
8339 if (OldRetTy != FT->getReturnType()) {
8340 if (Callee->isDeclaration() && !Caller->use_empty() &&
8341 // Conversion is ok if changing from pointer to int of same size.
8342 !(isa<PointerType>(FT->getReturnType()) &&
8343 TD->getIntPtrType() == OldRetTy))
8344 return false; // Cannot transform this return value.
8346 if (!Caller->use_empty() &&
8347 // void -> non-void is handled specially
8348 FT->getReturnType() != Type::VoidTy &&
8349 !CastInst::isCastable(FT->getReturnType(), OldRetTy))
8350 return false; // Cannot transform this return value.
8352 if (CallerPAL && !Caller->use_empty()) {
8353 uint16_t RAttrs = CallerPAL->getParamAttrs(0);
8354 if (RAttrs & ParamAttr::typeIncompatible(FT->getReturnType()))
8355 return false; // Attribute not compatible with transformed value.
8358 // If the callsite is an invoke instruction, and the return value is used by
8359 // a PHI node in a successor, we cannot change the return type of the call
8360 // because there is no place to put the cast instruction (without breaking
8361 // the critical edge). Bail out in this case.
8362 if (!Caller->use_empty())
8363 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller))
8364 for (Value::use_iterator UI = II->use_begin(), E = II->use_end();
8366 if (PHINode *PN = dyn_cast<PHINode>(*UI))
8367 if (PN->getParent() == II->getNormalDest() ||
8368 PN->getParent() == II->getUnwindDest())
8372 unsigned NumActualArgs = unsigned(CS.arg_end()-CS.arg_begin());
8373 unsigned NumCommonArgs = std::min(FT->getNumParams(), NumActualArgs);
8375 CallSite::arg_iterator AI = CS.arg_begin();
8376 for (unsigned i = 0, e = NumCommonArgs; i != e; ++i, ++AI) {
8377 const Type *ParamTy = FT->getParamType(i);
8378 const Type *ActTy = (*AI)->getType();
8380 if (!CastInst::isCastable(ActTy, ParamTy))
8381 return false; // Cannot transform this parameter value.
8384 uint16_t PAttrs = CallerPAL->getParamAttrs(i + 1);
8385 if (PAttrs & ParamAttr::typeIncompatible(ParamTy))
8386 return false; // Attribute not compatible with transformed value.
8389 ConstantInt *c = dyn_cast<ConstantInt>(*AI);
8390 // Some conversions are safe even if we do not have a body.
8391 // Either we can cast directly, or we can upconvert the argument
8392 bool isConvertible = ActTy == ParamTy ||
8393 (isa<PointerType>(ParamTy) && isa<PointerType>(ActTy)) ||
8394 (ParamTy->isInteger() && ActTy->isInteger() &&
8395 ParamTy->getPrimitiveSizeInBits() >= ActTy->getPrimitiveSizeInBits()) ||
8396 (c && ParamTy->getPrimitiveSizeInBits() >= ActTy->getPrimitiveSizeInBits()
8397 && c->getValue().isStrictlyPositive());
8398 if (Callee->isDeclaration() && !isConvertible) return false;
8401 if (FT->getNumParams() < NumActualArgs && !FT->isVarArg() &&
8402 Callee->isDeclaration())
8403 return false; // Do not delete arguments unless we have a function body...
8405 if (FT->getNumParams() < NumActualArgs && FT->isVarArg() && CallerPAL)
8406 // In this case we have more arguments than the new function type, but we
8407 // won't be dropping them. Check that these extra arguments have attributes
8408 // that are compatible with being a vararg call argument.
8409 for (unsigned i = CallerPAL->size(); i; --i) {
8410 if (CallerPAL->getParamIndex(i - 1) <= FT->getNumParams())
8412 uint16_t PAttrs = CallerPAL->getParamAttrsAtIndex(i - 1);
8413 if (PAttrs & ParamAttr::VarArgsIncompatible)
8417 // Okay, we decided that this is a safe thing to do: go ahead and start
8418 // inserting cast instructions as necessary...
8419 std::vector<Value*> Args;
8420 Args.reserve(NumActualArgs);
8421 ParamAttrsVector attrVec;
8422 attrVec.reserve(NumCommonArgs);
8424 // Get any return attributes.
8425 uint16_t RAttrs = CallerPAL ? CallerPAL->getParamAttrs(0) : 0;
8427 // If the return value is not being used, the type may not be compatible
8428 // with the existing attributes. Wipe out any problematic attributes.
8429 RAttrs &= ~ParamAttr::typeIncompatible(FT->getReturnType());
8431 // Add the new return attributes.
8433 attrVec.push_back(ParamAttrsWithIndex::get(0, RAttrs));
8435 AI = CS.arg_begin();
8436 for (unsigned i = 0; i != NumCommonArgs; ++i, ++AI) {
8437 const Type *ParamTy = FT->getParamType(i);
8438 if ((*AI)->getType() == ParamTy) {
8439 Args.push_back(*AI);
8441 Instruction::CastOps opcode = CastInst::getCastOpcode(*AI,
8442 false, ParamTy, false);
8443 CastInst *NewCast = CastInst::create(opcode, *AI, ParamTy, "tmp");
8444 Args.push_back(InsertNewInstBefore(NewCast, *Caller));
8447 // Add any parameter attributes.
8448 uint16_t PAttrs = CallerPAL ? CallerPAL->getParamAttrs(i + 1) : 0;
8450 attrVec.push_back(ParamAttrsWithIndex::get(i + 1, PAttrs));
8453 // If the function takes more arguments than the call was taking, add them
8455 for (unsigned i = NumCommonArgs; i != FT->getNumParams(); ++i)
8456 Args.push_back(Constant::getNullValue(FT->getParamType(i)));
8458 // If we are removing arguments to the function, emit an obnoxious warning...
8459 if (FT->getNumParams() < NumActualArgs)
8460 if (!FT->isVarArg()) {
8461 cerr << "WARNING: While resolving call to function '"
8462 << Callee->getName() << "' arguments were dropped!\n";
8464 // Add all of the arguments in their promoted form to the arg list...
8465 for (unsigned i = FT->getNumParams(); i != NumActualArgs; ++i, ++AI) {
8466 const Type *PTy = getPromotedType((*AI)->getType());
8467 if (PTy != (*AI)->getType()) {
8468 // Must promote to pass through va_arg area!
8469 Instruction::CastOps opcode = CastInst::getCastOpcode(*AI, false,
8471 Instruction *Cast = CastInst::create(opcode, *AI, PTy, "tmp");
8472 InsertNewInstBefore(Cast, *Caller);
8473 Args.push_back(Cast);
8475 Args.push_back(*AI);
8478 // Add any parameter attributes.
8479 uint16_t PAttrs = CallerPAL ? CallerPAL->getParamAttrs(i + 1) : 0;
8481 attrVec.push_back(ParamAttrsWithIndex::get(i + 1, PAttrs));
8485 if (FT->getReturnType() == Type::VoidTy)
8486 Caller->setName(""); // Void type should not have a name.
8488 const ParamAttrsList* NewCallerPAL = ParamAttrsList::get(attrVec);
8491 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
8492 NC = new InvokeInst(Callee, II->getNormalDest(), II->getUnwindDest(),
8493 Args.begin(), Args.end(), Caller->getName(), Caller);
8494 cast<InvokeInst>(NC)->setCallingConv(II->getCallingConv());
8495 cast<InvokeInst>(NC)->setParamAttrs(NewCallerPAL);
8497 NC = new CallInst(Callee, Args.begin(), Args.end(),
8498 Caller->getName(), Caller);
8499 CallInst *CI = cast<CallInst>(Caller);
8500 if (CI->isTailCall())
8501 cast<CallInst>(NC)->setTailCall();
8502 cast<CallInst>(NC)->setCallingConv(CI->getCallingConv());
8503 cast<CallInst>(NC)->setParamAttrs(NewCallerPAL);
8506 // Insert a cast of the return type as necessary.
8508 if (OldRetTy != NV->getType() && !Caller->use_empty()) {
8509 if (NV->getType() != Type::VoidTy) {
8510 Instruction::CastOps opcode = CastInst::getCastOpcode(NC, false,
8512 NV = NC = CastInst::create(opcode, NC, OldRetTy, "tmp");
8514 // If this is an invoke instruction, we should insert it after the first
8515 // non-phi, instruction in the normal successor block.
8516 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
8517 BasicBlock::iterator I = II->getNormalDest()->begin();
8518 while (isa<PHINode>(I)) ++I;
8519 InsertNewInstBefore(NC, *I);
8521 // Otherwise, it's a call, just insert cast right after the call instr
8522 InsertNewInstBefore(NC, *Caller);
8524 AddUsersToWorkList(*Caller);
8526 NV = UndefValue::get(Caller->getType());
8530 if (Caller->getType() != Type::VoidTy && !Caller->use_empty())
8531 Caller->replaceAllUsesWith(NV);
8532 Caller->eraseFromParent();
8533 RemoveFromWorkList(Caller);
8537 // transformCallThroughTrampoline - Turn a call to a function created by the
8538 // init_trampoline intrinsic into a direct call to the underlying function.
8540 Instruction *InstCombiner::transformCallThroughTrampoline(CallSite CS) {
8541 Value *Callee = CS.getCalledValue();
8542 const PointerType *PTy = cast<PointerType>(Callee->getType());
8543 const FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
8544 const ParamAttrsList *Attrs = CS.getParamAttrs();
8546 // If the call already has the 'nest' attribute somewhere then give up -
8547 // otherwise 'nest' would occur twice after splicing in the chain.
8548 if (Attrs && Attrs->hasAttrSomewhere(ParamAttr::Nest))
8551 IntrinsicInst *Tramp =
8552 cast<IntrinsicInst>(cast<BitCastInst>(Callee)->getOperand(0));
8555 cast<Function>(IntrinsicInst::StripPointerCasts(Tramp->getOperand(2)));
8556 const PointerType *NestFPTy = cast<PointerType>(NestF->getType());
8557 const FunctionType *NestFTy = cast<FunctionType>(NestFPTy->getElementType());
8559 if (const ParamAttrsList *NestAttrs = NestF->getParamAttrs()) {
8560 unsigned NestIdx = 1;
8561 const Type *NestTy = 0;
8562 uint16_t NestAttr = 0;
8564 // Look for a parameter marked with the 'nest' attribute.
8565 for (FunctionType::param_iterator I = NestFTy->param_begin(),
8566 E = NestFTy->param_end(); I != E; ++NestIdx, ++I)
8567 if (NestAttrs->paramHasAttr(NestIdx, ParamAttr::Nest)) {
8568 // Record the parameter type and any other attributes.
8570 NestAttr = NestAttrs->getParamAttrs(NestIdx);
8575 Instruction *Caller = CS.getInstruction();
8576 std::vector<Value*> NewArgs;
8577 NewArgs.reserve(unsigned(CS.arg_end()-CS.arg_begin())+1);
8579 ParamAttrsVector NewAttrs;
8580 NewAttrs.reserve(Attrs ? Attrs->size() + 1 : 1);
8582 // Insert the nest argument into the call argument list, which may
8583 // mean appending it. Likewise for attributes.
8585 // Add any function result attributes.
8586 uint16_t Attr = Attrs ? Attrs->getParamAttrs(0) : 0;
8588 NewAttrs.push_back (ParamAttrsWithIndex::get(0, Attr));
8592 CallSite::arg_iterator I = CS.arg_begin(), E = CS.arg_end();
8594 if (Idx == NestIdx) {
8595 // Add the chain argument and attributes.
8596 Value *NestVal = Tramp->getOperand(3);
8597 if (NestVal->getType() != NestTy)
8598 NestVal = new BitCastInst(NestVal, NestTy, "nest", Caller);
8599 NewArgs.push_back(NestVal);
8600 NewAttrs.push_back(ParamAttrsWithIndex::get(NestIdx, NestAttr));
8606 // Add the original argument and attributes.
8607 NewArgs.push_back(*I);
8608 Attr = Attrs ? Attrs->getParamAttrs(Idx) : 0;
8611 (ParamAttrsWithIndex::get(Idx + (Idx >= NestIdx), Attr));
8617 // The trampoline may have been bitcast to a bogus type (FTy).
8618 // Handle this by synthesizing a new function type, equal to FTy
8619 // with the chain parameter inserted.
8621 std::vector<const Type*> NewTypes;
8622 NewTypes.reserve(FTy->getNumParams()+1);
8624 // Insert the chain's type into the list of parameter types, which may
8625 // mean appending it.
8628 FunctionType::param_iterator I = FTy->param_begin(),
8629 E = FTy->param_end();
8633 // Add the chain's type.
8634 NewTypes.push_back(NestTy);
8639 // Add the original type.
8640 NewTypes.push_back(*I);
8646 // Replace the trampoline call with a direct call. Let the generic
8647 // code sort out any function type mismatches.
8648 FunctionType *NewFTy =
8649 FunctionType::get(FTy->getReturnType(), NewTypes, FTy->isVarArg());
8650 Constant *NewCallee = NestF->getType() == PointerType::getUnqual(NewFTy) ?
8651 NestF : ConstantExpr::getBitCast(NestF, PointerType::getUnqual(NewFTy));
8652 const ParamAttrsList *NewPAL = ParamAttrsList::get(NewAttrs);
8654 Instruction *NewCaller;
8655 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
8656 NewCaller = new InvokeInst(NewCallee,
8657 II->getNormalDest(), II->getUnwindDest(),
8658 NewArgs.begin(), NewArgs.end(),
8659 Caller->getName(), Caller);
8660 cast<InvokeInst>(NewCaller)->setCallingConv(II->getCallingConv());
8661 cast<InvokeInst>(NewCaller)->setParamAttrs(NewPAL);
8663 NewCaller = new CallInst(NewCallee, NewArgs.begin(), NewArgs.end(),
8664 Caller->getName(), Caller);
8665 if (cast<CallInst>(Caller)->isTailCall())
8666 cast<CallInst>(NewCaller)->setTailCall();
8667 cast<CallInst>(NewCaller)->
8668 setCallingConv(cast<CallInst>(Caller)->getCallingConv());
8669 cast<CallInst>(NewCaller)->setParamAttrs(NewPAL);
8671 if (Caller->getType() != Type::VoidTy && !Caller->use_empty())
8672 Caller->replaceAllUsesWith(NewCaller);
8673 Caller->eraseFromParent();
8674 RemoveFromWorkList(Caller);
8679 // Replace the trampoline call with a direct call. Since there is no 'nest'
8680 // parameter, there is no need to adjust the argument list. Let the generic
8681 // code sort out any function type mismatches.
8682 Constant *NewCallee =
8683 NestF->getType() == PTy ? NestF : ConstantExpr::getBitCast(NestF, PTy);
8684 CS.setCalledFunction(NewCallee);
8685 return CS.getInstruction();
8688 /// FoldPHIArgBinOpIntoPHI - If we have something like phi [add (a,b), add(c,d)]
8689 /// and if a/b/c/d and the add's all have a single use, turn this into two phi's
8690 /// and a single binop.
8691 Instruction *InstCombiner::FoldPHIArgBinOpIntoPHI(PHINode &PN) {
8692 Instruction *FirstInst = cast<Instruction>(PN.getIncomingValue(0));
8693 assert(isa<BinaryOperator>(FirstInst) || isa<GetElementPtrInst>(FirstInst) ||
8694 isa<CmpInst>(FirstInst));
8695 unsigned Opc = FirstInst->getOpcode();
8696 Value *LHSVal = FirstInst->getOperand(0);
8697 Value *RHSVal = FirstInst->getOperand(1);
8699 const Type *LHSType = LHSVal->getType();
8700 const Type *RHSType = RHSVal->getType();
8702 // Scan to see if all operands are the same opcode, all have one use, and all
8703 // kill their operands (i.e. the operands have one use).
8704 for (unsigned i = 0; i != PN.getNumIncomingValues(); ++i) {
8705 Instruction *I = dyn_cast<Instruction>(PN.getIncomingValue(i));
8706 if (!I || I->getOpcode() != Opc || !I->hasOneUse() ||
8707 // Verify type of the LHS matches so we don't fold cmp's of different
8708 // types or GEP's with different index types.
8709 I->getOperand(0)->getType() != LHSType ||
8710 I->getOperand(1)->getType() != RHSType)
8713 // If they are CmpInst instructions, check their predicates
8714 if (Opc == Instruction::ICmp || Opc == Instruction::FCmp)
8715 if (cast<CmpInst>(I)->getPredicate() !=
8716 cast<CmpInst>(FirstInst)->getPredicate())
8719 // Keep track of which operand needs a phi node.
8720 if (I->getOperand(0) != LHSVal) LHSVal = 0;
8721 if (I->getOperand(1) != RHSVal) RHSVal = 0;
8724 // Otherwise, this is safe to transform, determine if it is profitable.
8726 // If this is a GEP, and if the index (not the pointer) needs a PHI, bail out.
8727 // Indexes are often folded into load/store instructions, so we don't want to
8728 // hide them behind a phi.
8729 if (isa<GetElementPtrInst>(FirstInst) && RHSVal == 0)
8732 Value *InLHS = FirstInst->getOperand(0);
8733 Value *InRHS = FirstInst->getOperand(1);
8734 PHINode *NewLHS = 0, *NewRHS = 0;
8736 NewLHS = new PHINode(LHSType, FirstInst->getOperand(0)->getName()+".pn");
8737 NewLHS->reserveOperandSpace(PN.getNumOperands()/2);
8738 NewLHS->addIncoming(InLHS, PN.getIncomingBlock(0));
8739 InsertNewInstBefore(NewLHS, PN);
8744 NewRHS = new PHINode(RHSType, FirstInst->getOperand(1)->getName()+".pn");
8745 NewRHS->reserveOperandSpace(PN.getNumOperands()/2);
8746 NewRHS->addIncoming(InRHS, PN.getIncomingBlock(0));
8747 InsertNewInstBefore(NewRHS, PN);
8751 // Add all operands to the new PHIs.
8752 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
8754 Value *NewInLHS =cast<Instruction>(PN.getIncomingValue(i))->getOperand(0);
8755 NewLHS->addIncoming(NewInLHS, PN.getIncomingBlock(i));
8758 Value *NewInRHS =cast<Instruction>(PN.getIncomingValue(i))->getOperand(1);
8759 NewRHS->addIncoming(NewInRHS, PN.getIncomingBlock(i));
8763 if (BinaryOperator *BinOp = dyn_cast<BinaryOperator>(FirstInst))
8764 return BinaryOperator::create(BinOp->getOpcode(), LHSVal, RHSVal);
8765 else if (CmpInst *CIOp = dyn_cast<CmpInst>(FirstInst))
8766 return CmpInst::create(CIOp->getOpcode(), CIOp->getPredicate(), LHSVal,
8769 assert(isa<GetElementPtrInst>(FirstInst));
8770 return new GetElementPtrInst(LHSVal, RHSVal);
8774 /// isSafeToSinkLoad - Return true if we know that it is safe sink the load out
8775 /// of the block that defines it. This means that it must be obvious the value
8776 /// of the load is not changed from the point of the load to the end of the
8779 /// Finally, it is safe, but not profitable, to sink a load targetting a
8780 /// non-address-taken alloca. Doing so will cause us to not promote the alloca
8782 static bool isSafeToSinkLoad(LoadInst *L) {
8783 BasicBlock::iterator BBI = L, E = L->getParent()->end();
8785 for (++BBI; BBI != E; ++BBI)
8786 if (BBI->mayWriteToMemory())
8789 // Check for non-address taken alloca. If not address-taken already, it isn't
8790 // profitable to do this xform.
8791 if (AllocaInst *AI = dyn_cast<AllocaInst>(L->getOperand(0))) {
8792 bool isAddressTaken = false;
8793 for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end();
8795 if (isa<LoadInst>(UI)) continue;
8796 if (StoreInst *SI = dyn_cast<StoreInst>(*UI)) {
8797 // If storing TO the alloca, then the address isn't taken.
8798 if (SI->getOperand(1) == AI) continue;
8800 isAddressTaken = true;
8804 if (!isAddressTaken)
8812 // FoldPHIArgOpIntoPHI - If all operands to a PHI node are the same "unary"
8813 // operator and they all are only used by the PHI, PHI together their
8814 // inputs, and do the operation once, to the result of the PHI.
8815 Instruction *InstCombiner::FoldPHIArgOpIntoPHI(PHINode &PN) {
8816 Instruction *FirstInst = cast<Instruction>(PN.getIncomingValue(0));
8818 // Scan the instruction, looking for input operations that can be folded away.
8819 // If all input operands to the phi are the same instruction (e.g. a cast from
8820 // the same type or "+42") we can pull the operation through the PHI, reducing
8821 // code size and simplifying code.
8822 Constant *ConstantOp = 0;
8823 const Type *CastSrcTy = 0;
8824 bool isVolatile = false;
8825 if (isa<CastInst>(FirstInst)) {
8826 CastSrcTy = FirstInst->getOperand(0)->getType();
8827 } else if (isa<BinaryOperator>(FirstInst) || isa<CmpInst>(FirstInst)) {
8828 // Can fold binop, compare or shift here if the RHS is a constant,
8829 // otherwise call FoldPHIArgBinOpIntoPHI.
8830 ConstantOp = dyn_cast<Constant>(FirstInst->getOperand(1));
8831 if (ConstantOp == 0)
8832 return FoldPHIArgBinOpIntoPHI(PN);
8833 } else if (LoadInst *LI = dyn_cast<LoadInst>(FirstInst)) {
8834 isVolatile = LI->isVolatile();
8835 // We can't sink the load if the loaded value could be modified between the
8836 // load and the PHI.
8837 if (LI->getParent() != PN.getIncomingBlock(0) ||
8838 !isSafeToSinkLoad(LI))
8840 } else if (isa<GetElementPtrInst>(FirstInst)) {
8841 if (FirstInst->getNumOperands() == 2)
8842 return FoldPHIArgBinOpIntoPHI(PN);
8843 // Can't handle general GEPs yet.
8846 return 0; // Cannot fold this operation.
8849 // Check to see if all arguments are the same operation.
8850 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
8851 if (!isa<Instruction>(PN.getIncomingValue(i))) return 0;
8852 Instruction *I = cast<Instruction>(PN.getIncomingValue(i));
8853 if (!I->hasOneUse() || !I->isSameOperationAs(FirstInst))
8856 if (I->getOperand(0)->getType() != CastSrcTy)
8857 return 0; // Cast operation must match.
8858 } else if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
8859 // We can't sink the load if the loaded value could be modified between
8860 // the load and the PHI.
8861 if (LI->isVolatile() != isVolatile ||
8862 LI->getParent() != PN.getIncomingBlock(i) ||
8863 !isSafeToSinkLoad(LI))
8865 } else if (I->getOperand(1) != ConstantOp) {
8870 // Okay, they are all the same operation. Create a new PHI node of the
8871 // correct type, and PHI together all of the LHS's of the instructions.
8872 PHINode *NewPN = new PHINode(FirstInst->getOperand(0)->getType(),
8873 PN.getName()+".in");
8874 NewPN->reserveOperandSpace(PN.getNumOperands()/2);
8876 Value *InVal = FirstInst->getOperand(0);
8877 NewPN->addIncoming(InVal, PN.getIncomingBlock(0));
8879 // Add all operands to the new PHI.
8880 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
8881 Value *NewInVal = cast<Instruction>(PN.getIncomingValue(i))->getOperand(0);
8882 if (NewInVal != InVal)
8884 NewPN->addIncoming(NewInVal, PN.getIncomingBlock(i));
8889 // The new PHI unions all of the same values together. This is really
8890 // common, so we handle it intelligently here for compile-time speed.
8894 InsertNewInstBefore(NewPN, PN);
8898 // Insert and return the new operation.
8899 if (CastInst* FirstCI = dyn_cast<CastInst>(FirstInst))
8900 return CastInst::create(FirstCI->getOpcode(), PhiVal, PN.getType());
8901 else if (isa<LoadInst>(FirstInst))
8902 return new LoadInst(PhiVal, "", isVolatile);
8903 else if (BinaryOperator *BinOp = dyn_cast<BinaryOperator>(FirstInst))
8904 return BinaryOperator::create(BinOp->getOpcode(), PhiVal, ConstantOp);
8905 else if (CmpInst *CIOp = dyn_cast<CmpInst>(FirstInst))
8906 return CmpInst::create(CIOp->getOpcode(), CIOp->getPredicate(),
8907 PhiVal, ConstantOp);
8909 assert(0 && "Unknown operation");
8913 /// DeadPHICycle - Return true if this PHI node is only used by a PHI node cycle
8915 static bool DeadPHICycle(PHINode *PN,
8916 SmallPtrSet<PHINode*, 16> &PotentiallyDeadPHIs) {
8917 if (PN->use_empty()) return true;
8918 if (!PN->hasOneUse()) return false;
8920 // Remember this node, and if we find the cycle, return.
8921 if (!PotentiallyDeadPHIs.insert(PN))
8924 // Don't scan crazily complex things.
8925 if (PotentiallyDeadPHIs.size() == 16)
8928 if (PHINode *PU = dyn_cast<PHINode>(PN->use_back()))
8929 return DeadPHICycle(PU, PotentiallyDeadPHIs);
8934 /// PHIsEqualValue - Return true if this phi node is always equal to
8935 /// NonPhiInVal. This happens with mutually cyclic phi nodes like:
8936 /// z = some value; x = phi (y, z); y = phi (x, z)
8937 static bool PHIsEqualValue(PHINode *PN, Value *NonPhiInVal,
8938 SmallPtrSet<PHINode*, 16> &ValueEqualPHIs) {
8939 // See if we already saw this PHI node.
8940 if (!ValueEqualPHIs.insert(PN))
8943 // Don't scan crazily complex things.
8944 if (ValueEqualPHIs.size() == 16)
8947 // Scan the operands to see if they are either phi nodes or are equal to
8949 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
8950 Value *Op = PN->getIncomingValue(i);
8951 if (PHINode *OpPN = dyn_cast<PHINode>(Op)) {
8952 if (!PHIsEqualValue(OpPN, NonPhiInVal, ValueEqualPHIs))
8954 } else if (Op != NonPhiInVal)
8962 // PHINode simplification
8964 Instruction *InstCombiner::visitPHINode(PHINode &PN) {
8965 // If LCSSA is around, don't mess with Phi nodes
8966 if (MustPreserveLCSSA) return 0;
8968 if (Value *V = PN.hasConstantValue())
8969 return ReplaceInstUsesWith(PN, V);
8971 // If all PHI operands are the same operation, pull them through the PHI,
8972 // reducing code size.
8973 if (isa<Instruction>(PN.getIncomingValue(0)) &&
8974 PN.getIncomingValue(0)->hasOneUse())
8975 if (Instruction *Result = FoldPHIArgOpIntoPHI(PN))
8978 // If this is a trivial cycle in the PHI node graph, remove it. Basically, if
8979 // this PHI only has a single use (a PHI), and if that PHI only has one use (a
8980 // PHI)... break the cycle.
8981 if (PN.hasOneUse()) {
8982 Instruction *PHIUser = cast<Instruction>(PN.use_back());
8983 if (PHINode *PU = dyn_cast<PHINode>(PHIUser)) {
8984 SmallPtrSet<PHINode*, 16> PotentiallyDeadPHIs;
8985 PotentiallyDeadPHIs.insert(&PN);
8986 if (DeadPHICycle(PU, PotentiallyDeadPHIs))
8987 return ReplaceInstUsesWith(PN, UndefValue::get(PN.getType()));
8990 // If this phi has a single use, and if that use just computes a value for
8991 // the next iteration of a loop, delete the phi. This occurs with unused
8992 // induction variables, e.g. "for (int j = 0; ; ++j);". Detecting this
8993 // common case here is good because the only other things that catch this
8994 // are induction variable analysis (sometimes) and ADCE, which is only run
8996 if (PHIUser->hasOneUse() &&
8997 (isa<BinaryOperator>(PHIUser) || isa<GetElementPtrInst>(PHIUser)) &&
8998 PHIUser->use_back() == &PN) {
8999 return ReplaceInstUsesWith(PN, UndefValue::get(PN.getType()));
9003 // We sometimes end up with phi cycles that non-obviously end up being the
9004 // same value, for example:
9005 // z = some value; x = phi (y, z); y = phi (x, z)
9006 // where the phi nodes don't necessarily need to be in the same block. Do a
9007 // quick check to see if the PHI node only contains a single non-phi value, if
9008 // so, scan to see if the phi cycle is actually equal to that value.
9010 unsigned InValNo = 0, NumOperandVals = PN.getNumIncomingValues();
9011 // Scan for the first non-phi operand.
9012 while (InValNo != NumOperandVals &&
9013 isa<PHINode>(PN.getIncomingValue(InValNo)))
9016 if (InValNo != NumOperandVals) {
9017 Value *NonPhiInVal = PN.getOperand(InValNo);
9019 // Scan the rest of the operands to see if there are any conflicts, if so
9020 // there is no need to recursively scan other phis.
9021 for (++InValNo; InValNo != NumOperandVals; ++InValNo) {
9022 Value *OpVal = PN.getIncomingValue(InValNo);
9023 if (OpVal != NonPhiInVal && !isa<PHINode>(OpVal))
9027 // If we scanned over all operands, then we have one unique value plus
9028 // phi values. Scan PHI nodes to see if they all merge in each other or
9030 if (InValNo == NumOperandVals) {
9031 SmallPtrSet<PHINode*, 16> ValueEqualPHIs;
9032 if (PHIsEqualValue(&PN, NonPhiInVal, ValueEqualPHIs))
9033 return ReplaceInstUsesWith(PN, NonPhiInVal);
9040 static Value *InsertCastToIntPtrTy(Value *V, const Type *DTy,
9041 Instruction *InsertPoint,
9043 unsigned PtrSize = DTy->getPrimitiveSizeInBits();
9044 unsigned VTySize = V->getType()->getPrimitiveSizeInBits();
9045 // We must cast correctly to the pointer type. Ensure that we
9046 // sign extend the integer value if it is smaller as this is
9047 // used for address computation.
9048 Instruction::CastOps opcode =
9049 (VTySize < PtrSize ? Instruction::SExt :
9050 (VTySize == PtrSize ? Instruction::BitCast : Instruction::Trunc));
9051 return IC->InsertCastBefore(opcode, V, DTy, *InsertPoint);
9055 Instruction *InstCombiner::visitGetElementPtrInst(GetElementPtrInst &GEP) {
9056 Value *PtrOp = GEP.getOperand(0);
9057 // Is it 'getelementptr %P, i32 0' or 'getelementptr %P'
9058 // If so, eliminate the noop.
9059 if (GEP.getNumOperands() == 1)
9060 return ReplaceInstUsesWith(GEP, PtrOp);
9062 if (isa<UndefValue>(GEP.getOperand(0)))
9063 return ReplaceInstUsesWith(GEP, UndefValue::get(GEP.getType()));
9065 bool HasZeroPointerIndex = false;
9066 if (Constant *C = dyn_cast<Constant>(GEP.getOperand(1)))
9067 HasZeroPointerIndex = C->isNullValue();
9069 if (GEP.getNumOperands() == 2 && HasZeroPointerIndex)
9070 return ReplaceInstUsesWith(GEP, PtrOp);
9072 // Eliminate unneeded casts for indices.
9073 bool MadeChange = false;
9075 gep_type_iterator GTI = gep_type_begin(GEP);
9076 for (unsigned i = 1, e = GEP.getNumOperands(); i != e; ++i, ++GTI) {
9077 if (isa<SequentialType>(*GTI)) {
9078 if (CastInst *CI = dyn_cast<CastInst>(GEP.getOperand(i))) {
9079 if (CI->getOpcode() == Instruction::ZExt ||
9080 CI->getOpcode() == Instruction::SExt) {
9081 const Type *SrcTy = CI->getOperand(0)->getType();
9082 // We can eliminate a cast from i32 to i64 iff the target
9083 // is a 32-bit pointer target.
9084 if (SrcTy->getPrimitiveSizeInBits() >= TD->getPointerSizeInBits()) {
9086 GEP.setOperand(i, CI->getOperand(0));
9090 // If we are using a wider index than needed for this platform, shrink it
9091 // to what we need. If the incoming value needs a cast instruction,
9092 // insert it. This explicit cast can make subsequent optimizations more
9094 Value *Op = GEP.getOperand(i);
9095 if (TD->getTypeSizeInBits(Op->getType()) > TD->getPointerSizeInBits())
9096 if (Constant *C = dyn_cast<Constant>(Op)) {
9097 GEP.setOperand(i, ConstantExpr::getTrunc(C, TD->getIntPtrType()));
9100 Op = InsertCastBefore(Instruction::Trunc, Op, TD->getIntPtrType(),
9102 GEP.setOperand(i, Op);
9107 if (MadeChange) return &GEP;
9109 // If this GEP instruction doesn't move the pointer, and if the input operand
9110 // is a bitcast of another pointer, just replace the GEP with a bitcast of the
9111 // real input to the dest type.
9112 if (GEP.hasAllZeroIndices()) {
9113 if (BitCastInst *BCI = dyn_cast<BitCastInst>(GEP.getOperand(0))) {
9114 // If the bitcast is of an allocation, and the allocation will be
9115 // converted to match the type of the cast, don't touch this.
9116 if (isa<AllocationInst>(BCI->getOperand(0))) {
9117 // See if the bitcast simplifies, if so, don't nuke this GEP yet.
9118 if (Instruction *I = visitBitCast(*BCI)) {
9121 BCI->getParent()->getInstList().insert(BCI, I);
9122 ReplaceInstUsesWith(*BCI, I);
9127 return new BitCastInst(BCI->getOperand(0), GEP.getType());
9131 // Combine Indices - If the source pointer to this getelementptr instruction
9132 // is a getelementptr instruction, combine the indices of the two
9133 // getelementptr instructions into a single instruction.
9135 SmallVector<Value*, 8> SrcGEPOperands;
9136 if (User *Src = dyn_castGetElementPtr(PtrOp))
9137 SrcGEPOperands.append(Src->op_begin(), Src->op_end());
9139 if (!SrcGEPOperands.empty()) {
9140 // Note that if our source is a gep chain itself that we wait for that
9141 // chain to be resolved before we perform this transformation. This
9142 // avoids us creating a TON of code in some cases.
9144 if (isa<GetElementPtrInst>(SrcGEPOperands[0]) &&
9145 cast<Instruction>(SrcGEPOperands[0])->getNumOperands() == 2)
9146 return 0; // Wait until our source is folded to completion.
9148 SmallVector<Value*, 8> Indices;
9150 // Find out whether the last index in the source GEP is a sequential idx.
9151 bool EndsWithSequential = false;
9152 for (gep_type_iterator I = gep_type_begin(*cast<User>(PtrOp)),
9153 E = gep_type_end(*cast<User>(PtrOp)); I != E; ++I)
9154 EndsWithSequential = !isa<StructType>(*I);
9156 // Can we combine the two pointer arithmetics offsets?
9157 if (EndsWithSequential) {
9158 // Replace: gep (gep %P, long B), long A, ...
9159 // With: T = long A+B; gep %P, T, ...
9161 Value *Sum, *SO1 = SrcGEPOperands.back(), *GO1 = GEP.getOperand(1);
9162 if (SO1 == Constant::getNullValue(SO1->getType())) {
9164 } else if (GO1 == Constant::getNullValue(GO1->getType())) {
9167 // If they aren't the same type, convert both to an integer of the
9168 // target's pointer size.
9169 if (SO1->getType() != GO1->getType()) {
9170 if (Constant *SO1C = dyn_cast<Constant>(SO1)) {
9171 SO1 = ConstantExpr::getIntegerCast(SO1C, GO1->getType(), true);
9172 } else if (Constant *GO1C = dyn_cast<Constant>(GO1)) {
9173 GO1 = ConstantExpr::getIntegerCast(GO1C, SO1->getType(), true);
9175 unsigned PS = TD->getPointerSizeInBits();
9176 if (TD->getTypeSizeInBits(SO1->getType()) == PS) {
9177 // Convert GO1 to SO1's type.
9178 GO1 = InsertCastToIntPtrTy(GO1, SO1->getType(), &GEP, this);
9180 } else if (TD->getTypeSizeInBits(GO1->getType()) == PS) {
9181 // Convert SO1 to GO1's type.
9182 SO1 = InsertCastToIntPtrTy(SO1, GO1->getType(), &GEP, this);
9184 const Type *PT = TD->getIntPtrType();
9185 SO1 = InsertCastToIntPtrTy(SO1, PT, &GEP, this);
9186 GO1 = InsertCastToIntPtrTy(GO1, PT, &GEP, this);
9190 if (isa<Constant>(SO1) && isa<Constant>(GO1))
9191 Sum = ConstantExpr::getAdd(cast<Constant>(SO1), cast<Constant>(GO1));
9193 Sum = BinaryOperator::createAdd(SO1, GO1, PtrOp->getName()+".sum");
9194 InsertNewInstBefore(cast<Instruction>(Sum), GEP);
9198 // Recycle the GEP we already have if possible.
9199 if (SrcGEPOperands.size() == 2) {
9200 GEP.setOperand(0, SrcGEPOperands[0]);
9201 GEP.setOperand(1, Sum);
9204 Indices.insert(Indices.end(), SrcGEPOperands.begin()+1,
9205 SrcGEPOperands.end()-1);
9206 Indices.push_back(Sum);
9207 Indices.insert(Indices.end(), GEP.op_begin()+2, GEP.op_end());
9209 } else if (isa<Constant>(*GEP.idx_begin()) &&
9210 cast<Constant>(*GEP.idx_begin())->isNullValue() &&
9211 SrcGEPOperands.size() != 1) {
9212 // Otherwise we can do the fold if the first index of the GEP is a zero
9213 Indices.insert(Indices.end(), SrcGEPOperands.begin()+1,
9214 SrcGEPOperands.end());
9215 Indices.insert(Indices.end(), GEP.idx_begin()+1, GEP.idx_end());
9218 if (!Indices.empty())
9219 return new GetElementPtrInst(SrcGEPOperands[0], Indices.begin(),
9220 Indices.end(), GEP.getName());
9222 } else if (GlobalValue *GV = dyn_cast<GlobalValue>(PtrOp)) {
9223 // GEP of global variable. If all of the indices for this GEP are
9224 // constants, we can promote this to a constexpr instead of an instruction.
9226 // Scan for nonconstants...
9227 SmallVector<Constant*, 8> Indices;
9228 User::op_iterator I = GEP.idx_begin(), E = GEP.idx_end();
9229 for (; I != E && isa<Constant>(*I); ++I)
9230 Indices.push_back(cast<Constant>(*I));
9232 if (I == E) { // If they are all constants...
9233 Constant *CE = ConstantExpr::getGetElementPtr(GV,
9234 &Indices[0],Indices.size());
9236 // Replace all uses of the GEP with the new constexpr...
9237 return ReplaceInstUsesWith(GEP, CE);
9239 } else if (Value *X = getBitCastOperand(PtrOp)) { // Is the operand a cast?
9240 if (!isa<PointerType>(X->getType())) {
9241 // Not interesting. Source pointer must be a cast from pointer.
9242 } else if (HasZeroPointerIndex) {
9243 // transform: GEP (bitcast [10 x i8]* X to [0 x i8]*), i32 0, ...
9244 // into : GEP [10 x i8]* X, i32 0, ...
9246 // This occurs when the program declares an array extern like "int X[];"
9248 const PointerType *CPTy = cast<PointerType>(PtrOp->getType());
9249 const PointerType *XTy = cast<PointerType>(X->getType());
9250 if (const ArrayType *XATy =
9251 dyn_cast<ArrayType>(XTy->getElementType()))
9252 if (const ArrayType *CATy =
9253 dyn_cast<ArrayType>(CPTy->getElementType()))
9254 if (CATy->getElementType() == XATy->getElementType()) {
9255 // At this point, we know that the cast source type is a pointer
9256 // to an array of the same type as the destination pointer
9257 // array. Because the array type is never stepped over (there
9258 // is a leading zero) we can fold the cast into this GEP.
9259 GEP.setOperand(0, X);
9262 } else if (GEP.getNumOperands() == 2) {
9263 // Transform things like:
9264 // %t = getelementptr i32* bitcast ([2 x i32]* %str to i32*), i32 %V
9265 // into: %t1 = getelementptr [2 x i32]* %str, i32 0, i32 %V; bitcast
9266 const Type *SrcElTy = cast<PointerType>(X->getType())->getElementType();
9267 const Type *ResElTy=cast<PointerType>(PtrOp->getType())->getElementType();
9268 if (isa<ArrayType>(SrcElTy) &&
9269 TD->getABITypeSize(cast<ArrayType>(SrcElTy)->getElementType()) ==
9270 TD->getABITypeSize(ResElTy)) {
9272 Idx[0] = Constant::getNullValue(Type::Int32Ty);
9273 Idx[1] = GEP.getOperand(1);
9274 Value *V = InsertNewInstBefore(
9275 new GetElementPtrInst(X, Idx, Idx + 2, GEP.getName()), GEP);
9276 // V and GEP are both pointer types --> BitCast
9277 return new BitCastInst(V, GEP.getType());
9280 // Transform things like:
9281 // getelementptr i8* bitcast ([100 x double]* X to i8*), i32 %tmp
9282 // (where tmp = 8*tmp2) into:
9283 // getelementptr [100 x double]* %arr, i32 0, i32 %tmp2; bitcast
9285 if (isa<ArrayType>(SrcElTy) && ResElTy == Type::Int8Ty) {
9286 uint64_t ArrayEltSize =
9287 TD->getABITypeSize(cast<ArrayType>(SrcElTy)->getElementType());
9289 // Check to see if "tmp" is a scale by a multiple of ArrayEltSize. We
9290 // allow either a mul, shift, or constant here.
9292 ConstantInt *Scale = 0;
9293 if (ArrayEltSize == 1) {
9294 NewIdx = GEP.getOperand(1);
9295 Scale = ConstantInt::get(NewIdx->getType(), 1);
9296 } else if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP.getOperand(1))) {
9297 NewIdx = ConstantInt::get(CI->getType(), 1);
9299 } else if (Instruction *Inst =dyn_cast<Instruction>(GEP.getOperand(1))){
9300 if (Inst->getOpcode() == Instruction::Shl &&
9301 isa<ConstantInt>(Inst->getOperand(1))) {
9302 ConstantInt *ShAmt = cast<ConstantInt>(Inst->getOperand(1));
9303 uint32_t ShAmtVal = ShAmt->getLimitedValue(64);
9304 Scale = ConstantInt::get(Inst->getType(), 1ULL << ShAmtVal);
9305 NewIdx = Inst->getOperand(0);
9306 } else if (Inst->getOpcode() == Instruction::Mul &&
9307 isa<ConstantInt>(Inst->getOperand(1))) {
9308 Scale = cast<ConstantInt>(Inst->getOperand(1));
9309 NewIdx = Inst->getOperand(0);
9313 // If the index will be to exactly the right offset with the scale taken
9314 // out, perform the transformation. Note, we don't know whether Scale is
9315 // signed or not. We'll use unsigned version of division/modulo
9316 // operation after making sure Scale doesn't have the sign bit set.
9317 if (Scale && Scale->getSExtValue() >= 0LL &&
9318 Scale->getZExtValue() % ArrayEltSize == 0) {
9319 Scale = ConstantInt::get(Scale->getType(),
9320 Scale->getZExtValue() / ArrayEltSize);
9321 if (Scale->getZExtValue() != 1) {
9322 Constant *C = ConstantExpr::getIntegerCast(Scale, NewIdx->getType(),
9324 Instruction *Sc = BinaryOperator::createMul(NewIdx, C, "idxscale");
9325 NewIdx = InsertNewInstBefore(Sc, GEP);
9328 // Insert the new GEP instruction.
9330 Idx[0] = Constant::getNullValue(Type::Int32Ty);
9332 Instruction *NewGEP =
9333 new GetElementPtrInst(X, Idx, Idx + 2, GEP.getName());
9334 NewGEP = InsertNewInstBefore(NewGEP, GEP);
9335 // The NewGEP must be pointer typed, so must the old one -> BitCast
9336 return new BitCastInst(NewGEP, GEP.getType());
9345 Instruction *InstCombiner::visitAllocationInst(AllocationInst &AI) {
9346 // Convert: malloc Ty, C - where C is a constant != 1 into: malloc [C x Ty], 1
9347 if (AI.isArrayAllocation()) // Check C != 1
9348 if (const ConstantInt *C = dyn_cast<ConstantInt>(AI.getArraySize())) {
9350 ArrayType::get(AI.getAllocatedType(), C->getZExtValue());
9351 AllocationInst *New = 0;
9353 // Create and insert the replacement instruction...
9354 if (isa<MallocInst>(AI))
9355 New = new MallocInst(NewTy, 0, AI.getAlignment(), AI.getName());
9357 assert(isa<AllocaInst>(AI) && "Unknown type of allocation inst!");
9358 New = new AllocaInst(NewTy, 0, AI.getAlignment(), AI.getName());
9361 InsertNewInstBefore(New, AI);
9363 // Scan to the end of the allocation instructions, to skip over a block of
9364 // allocas if possible...
9366 BasicBlock::iterator It = New;
9367 while (isa<AllocationInst>(*It)) ++It;
9369 // Now that I is pointing to the first non-allocation-inst in the block,
9370 // insert our getelementptr instruction...
9372 Value *NullIdx = Constant::getNullValue(Type::Int32Ty);
9376 Value *V = new GetElementPtrInst(New, Idx, Idx + 2,
9377 New->getName()+".sub", It);
9379 // Now make everything use the getelementptr instead of the original
9381 return ReplaceInstUsesWith(AI, V);
9382 } else if (isa<UndefValue>(AI.getArraySize())) {
9383 return ReplaceInstUsesWith(AI, Constant::getNullValue(AI.getType()));
9386 // If alloca'ing a zero byte object, replace the alloca with a null pointer.
9387 // Note that we only do this for alloca's, because malloc should allocate and
9388 // return a unique pointer, even for a zero byte allocation.
9389 if (isa<AllocaInst>(AI) && AI.getAllocatedType()->isSized() &&
9390 TD->getABITypeSize(AI.getAllocatedType()) == 0)
9391 return ReplaceInstUsesWith(AI, Constant::getNullValue(AI.getType()));
9396 Instruction *InstCombiner::visitFreeInst(FreeInst &FI) {
9397 Value *Op = FI.getOperand(0);
9399 // free undef -> unreachable.
9400 if (isa<UndefValue>(Op)) {
9401 // Insert a new store to null because we cannot modify the CFG here.
9402 new StoreInst(ConstantInt::getTrue(),
9403 UndefValue::get(PointerType::getUnqual(Type::Int1Ty)), &FI);
9404 return EraseInstFromFunction(FI);
9407 // If we have 'free null' delete the instruction. This can happen in stl code
9408 // when lots of inlining happens.
9409 if (isa<ConstantPointerNull>(Op))
9410 return EraseInstFromFunction(FI);
9412 // Change free <ty>* (cast <ty2>* X to <ty>*) into free <ty2>* X
9413 if (BitCastInst *CI = dyn_cast<BitCastInst>(Op)) {
9414 FI.setOperand(0, CI->getOperand(0));
9418 // Change free (gep X, 0,0,0,0) into free(X)
9419 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(Op)) {
9420 if (GEPI->hasAllZeroIndices()) {
9421 AddToWorkList(GEPI);
9422 FI.setOperand(0, GEPI->getOperand(0));
9427 // Change free(malloc) into nothing, if the malloc has a single use.
9428 if (MallocInst *MI = dyn_cast<MallocInst>(Op))
9429 if (MI->hasOneUse()) {
9430 EraseInstFromFunction(FI);
9431 return EraseInstFromFunction(*MI);
9438 /// InstCombineLoadCast - Fold 'load (cast P)' -> cast (load P)' when possible.
9439 static Instruction *InstCombineLoadCast(InstCombiner &IC, LoadInst &LI,
9440 const TargetData *TD) {
9441 User *CI = cast<User>(LI.getOperand(0));
9442 Value *CastOp = CI->getOperand(0);
9444 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(CI)) {
9445 // Instead of loading constant c string, use corresponding integer value
9446 // directly if string length is small enough.
9447 const std::string &Str = CE->getOperand(0)->getStringValue();
9449 unsigned len = Str.length();
9450 const Type *Ty = cast<PointerType>(CE->getType())->getElementType();
9451 unsigned numBits = Ty->getPrimitiveSizeInBits();
9452 // Replace LI with immediate integer store.
9453 if ((numBits >> 3) == len + 1) {
9454 APInt StrVal(numBits, 0);
9455 APInt SingleChar(numBits, 0);
9456 if (TD->isLittleEndian()) {
9457 for (signed i = len-1; i >= 0; i--) {
9458 SingleChar = (uint64_t) Str[i];
9459 StrVal = (StrVal << 8) | SingleChar;
9462 for (unsigned i = 0; i < len; i++) {
9463 SingleChar = (uint64_t) Str[i];
9464 StrVal = (StrVal << 8) | SingleChar;
9466 // Append NULL at the end.
9468 StrVal = (StrVal << 8) | SingleChar;
9470 Value *NL = ConstantInt::get(StrVal);
9471 return IC.ReplaceInstUsesWith(LI, NL);
9476 const Type *DestPTy = cast<PointerType>(CI->getType())->getElementType();
9477 if (const PointerType *SrcTy = dyn_cast<PointerType>(CastOp->getType())) {
9478 const Type *SrcPTy = SrcTy->getElementType();
9480 if (DestPTy->isInteger() || isa<PointerType>(DestPTy) ||
9481 isa<VectorType>(DestPTy)) {
9482 // If the source is an array, the code below will not succeed. Check to
9483 // see if a trivial 'gep P, 0, 0' will help matters. Only do this for
9485 if (const ArrayType *ASrcTy = dyn_cast<ArrayType>(SrcPTy))
9486 if (Constant *CSrc = dyn_cast<Constant>(CastOp))
9487 if (ASrcTy->getNumElements() != 0) {
9489 Idxs[0] = Idxs[1] = Constant::getNullValue(Type::Int32Ty);
9490 CastOp = ConstantExpr::getGetElementPtr(CSrc, Idxs, 2);
9491 SrcTy = cast<PointerType>(CastOp->getType());
9492 SrcPTy = SrcTy->getElementType();
9495 if ((SrcPTy->isInteger() || isa<PointerType>(SrcPTy) ||
9496 isa<VectorType>(SrcPTy)) &&
9497 // Do not allow turning this into a load of an integer, which is then
9498 // casted to a pointer, this pessimizes pointer analysis a lot.
9499 (isa<PointerType>(SrcPTy) == isa<PointerType>(LI.getType())) &&
9500 IC.getTargetData().getTypeSizeInBits(SrcPTy) ==
9501 IC.getTargetData().getTypeSizeInBits(DestPTy)) {
9503 // Okay, we are casting from one integer or pointer type to another of
9504 // the same size. Instead of casting the pointer before the load, cast
9505 // the result of the loaded value.
9506 Value *NewLoad = IC.InsertNewInstBefore(new LoadInst(CastOp,
9508 LI.isVolatile()),LI);
9509 // Now cast the result of the load.
9510 return new BitCastInst(NewLoad, LI.getType());
9517 /// isSafeToLoadUnconditionally - Return true if we know that executing a load
9518 /// from this value cannot trap. If it is not obviously safe to load from the
9519 /// specified pointer, we do a quick local scan of the basic block containing
9520 /// ScanFrom, to determine if the address is already accessed.
9521 static bool isSafeToLoadUnconditionally(Value *V, Instruction *ScanFrom) {
9522 // If it is an alloca it is always safe to load from.
9523 if (isa<AllocaInst>(V)) return true;
9525 // If it is a global variable it is mostly safe to load from.
9526 if (const GlobalValue *GV = dyn_cast<GlobalVariable>(V))
9527 // Don't try to evaluate aliases. External weak GV can be null.
9528 return !isa<GlobalAlias>(GV) && !GV->hasExternalWeakLinkage();
9530 // Otherwise, be a little bit agressive by scanning the local block where we
9531 // want to check to see if the pointer is already being loaded or stored
9532 // from/to. If so, the previous load or store would have already trapped,
9533 // so there is no harm doing an extra load (also, CSE will later eliminate
9534 // the load entirely).
9535 BasicBlock::iterator BBI = ScanFrom, E = ScanFrom->getParent()->begin();
9540 if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) {
9541 if (LI->getOperand(0) == V) return true;
9542 } else if (StoreInst *SI = dyn_cast<StoreInst>(BBI))
9543 if (SI->getOperand(1) == V) return true;
9549 /// GetUnderlyingObject - Trace through a series of getelementptrs and bitcasts
9550 /// until we find the underlying object a pointer is referring to or something
9551 /// we don't understand. Note that the returned pointer may be offset from the
9552 /// input, because we ignore GEP indices.
9553 static Value *GetUnderlyingObject(Value *Ptr) {
9555 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ptr)) {
9556 if (CE->getOpcode() == Instruction::BitCast ||
9557 CE->getOpcode() == Instruction::GetElementPtr)
9558 Ptr = CE->getOperand(0);
9561 } else if (BitCastInst *BCI = dyn_cast<BitCastInst>(Ptr)) {
9562 Ptr = BCI->getOperand(0);
9563 } else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr)) {
9564 Ptr = GEP->getOperand(0);
9571 Instruction *InstCombiner::visitLoadInst(LoadInst &LI) {
9572 Value *Op = LI.getOperand(0);
9574 // Attempt to improve the alignment.
9575 unsigned KnownAlign = GetOrEnforceKnownAlignment(Op, TD);
9576 if (KnownAlign > LI.getAlignment())
9577 LI.setAlignment(KnownAlign);
9579 // load (cast X) --> cast (load X) iff safe
9580 if (isa<CastInst>(Op))
9581 if (Instruction *Res = InstCombineLoadCast(*this, LI, TD))
9584 // None of the following transforms are legal for volatile loads.
9585 if (LI.isVolatile()) return 0;
9587 if (&LI.getParent()->front() != &LI) {
9588 BasicBlock::iterator BBI = &LI; --BBI;
9589 // If the instruction immediately before this is a store to the same
9590 // address, do a simple form of store->load forwarding.
9591 if (StoreInst *SI = dyn_cast<StoreInst>(BBI))
9592 if (SI->getOperand(1) == LI.getOperand(0))
9593 return ReplaceInstUsesWith(LI, SI->getOperand(0));
9594 if (LoadInst *LIB = dyn_cast<LoadInst>(BBI))
9595 if (LIB->getOperand(0) == LI.getOperand(0))
9596 return ReplaceInstUsesWith(LI, LIB);
9599 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(Op)) {
9600 const Value *GEPI0 = GEPI->getOperand(0);
9601 // TODO: Consider a target hook for valid address spaces for this xform.
9602 if (isa<ConstantPointerNull>(GEPI0) &&
9603 cast<PointerType>(GEPI0->getType())->getAddressSpace() == 0) {
9604 // Insert a new store to null instruction before the load to indicate
9605 // that this code is not reachable. We do this instead of inserting
9606 // an unreachable instruction directly because we cannot modify the
9608 new StoreInst(UndefValue::get(LI.getType()),
9609 Constant::getNullValue(Op->getType()), &LI);
9610 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
9614 if (Constant *C = dyn_cast<Constant>(Op)) {
9615 // load null/undef -> undef
9616 // TODO: Consider a target hook for valid address spaces for this xform.
9617 if (isa<UndefValue>(C) || (C->isNullValue() &&
9618 cast<PointerType>(Op->getType())->getAddressSpace() == 0)) {
9619 // Insert a new store to null instruction before the load to indicate that
9620 // this code is not reachable. We do this instead of inserting an
9621 // unreachable instruction directly because we cannot modify the CFG.
9622 new StoreInst(UndefValue::get(LI.getType()),
9623 Constant::getNullValue(Op->getType()), &LI);
9624 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
9627 // Instcombine load (constant global) into the value loaded.
9628 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Op))
9629 if (GV->isConstant() && !GV->isDeclaration())
9630 return ReplaceInstUsesWith(LI, GV->getInitializer());
9632 // Instcombine load (constantexpr_GEP global, 0, ...) into the value loaded.
9633 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Op))
9634 if (CE->getOpcode() == Instruction::GetElementPtr) {
9635 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(CE->getOperand(0)))
9636 if (GV->isConstant() && !GV->isDeclaration())
9638 ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(), CE))
9639 return ReplaceInstUsesWith(LI, V);
9640 if (CE->getOperand(0)->isNullValue()) {
9641 // Insert a new store to null instruction before the load to indicate
9642 // that this code is not reachable. We do this instead of inserting
9643 // an unreachable instruction directly because we cannot modify the
9645 new StoreInst(UndefValue::get(LI.getType()),
9646 Constant::getNullValue(Op->getType()), &LI);
9647 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
9650 } else if (CE->isCast()) {
9651 if (Instruction *Res = InstCombineLoadCast(*this, LI, TD))
9656 // If this load comes from anywhere in a constant global, and if the global
9657 // is all undef or zero, we know what it loads.
9658 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(GetUnderlyingObject(Op))) {
9659 if (GV->isConstant() && GV->hasInitializer()) {
9660 if (GV->getInitializer()->isNullValue())
9661 return ReplaceInstUsesWith(LI, Constant::getNullValue(LI.getType()));
9662 else if (isa<UndefValue>(GV->getInitializer()))
9663 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
9667 if (Op->hasOneUse()) {
9668 // Change select and PHI nodes to select values instead of addresses: this
9669 // helps alias analysis out a lot, allows many others simplifications, and
9670 // exposes redundancy in the code.
9672 // Note that we cannot do the transformation unless we know that the
9673 // introduced loads cannot trap! Something like this is valid as long as
9674 // the condition is always false: load (select bool %C, int* null, int* %G),
9675 // but it would not be valid if we transformed it to load from null
9678 if (SelectInst *SI = dyn_cast<SelectInst>(Op)) {
9679 // load (select (Cond, &V1, &V2)) --> select(Cond, load &V1, load &V2).
9680 if (isSafeToLoadUnconditionally(SI->getOperand(1), SI) &&
9681 isSafeToLoadUnconditionally(SI->getOperand(2), SI)) {
9682 Value *V1 = InsertNewInstBefore(new LoadInst(SI->getOperand(1),
9683 SI->getOperand(1)->getName()+".val"), LI);
9684 Value *V2 = InsertNewInstBefore(new LoadInst(SI->getOperand(2),
9685 SI->getOperand(2)->getName()+".val"), LI);
9686 return new SelectInst(SI->getCondition(), V1, V2);
9689 // load (select (cond, null, P)) -> load P
9690 if (Constant *C = dyn_cast<Constant>(SI->getOperand(1)))
9691 if (C->isNullValue()) {
9692 LI.setOperand(0, SI->getOperand(2));
9696 // load (select (cond, P, null)) -> load P
9697 if (Constant *C = dyn_cast<Constant>(SI->getOperand(2)))
9698 if (C->isNullValue()) {
9699 LI.setOperand(0, SI->getOperand(1));
9707 /// InstCombineStoreToCast - Fold store V, (cast P) -> store (cast V), P
9709 static Instruction *InstCombineStoreToCast(InstCombiner &IC, StoreInst &SI) {
9710 User *CI = cast<User>(SI.getOperand(1));
9711 Value *CastOp = CI->getOperand(0);
9713 const Type *DestPTy = cast<PointerType>(CI->getType())->getElementType();
9714 if (const PointerType *SrcTy = dyn_cast<PointerType>(CastOp->getType())) {
9715 const Type *SrcPTy = SrcTy->getElementType();
9717 if (DestPTy->isInteger() || isa<PointerType>(DestPTy)) {
9718 // If the source is an array, the code below will not succeed. Check to
9719 // see if a trivial 'gep P, 0, 0' will help matters. Only do this for
9721 if (const ArrayType *ASrcTy = dyn_cast<ArrayType>(SrcPTy))
9722 if (Constant *CSrc = dyn_cast<Constant>(CastOp))
9723 if (ASrcTy->getNumElements() != 0) {
9725 Idxs[0] = Idxs[1] = Constant::getNullValue(Type::Int32Ty);
9726 CastOp = ConstantExpr::getGetElementPtr(CSrc, Idxs, 2);
9727 SrcTy = cast<PointerType>(CastOp->getType());
9728 SrcPTy = SrcTy->getElementType();
9731 if ((SrcPTy->isInteger() || isa<PointerType>(SrcPTy)) &&
9732 IC.getTargetData().getTypeSizeInBits(SrcPTy) ==
9733 IC.getTargetData().getTypeSizeInBits(DestPTy)) {
9735 // Okay, we are casting from one integer or pointer type to another of
9736 // the same size. Instead of casting the pointer before
9737 // the store, cast the value to be stored.
9739 Value *SIOp0 = SI.getOperand(0);
9740 Instruction::CastOps opcode = Instruction::BitCast;
9741 const Type* CastSrcTy = SIOp0->getType();
9742 const Type* CastDstTy = SrcPTy;
9743 if (isa<PointerType>(CastDstTy)) {
9744 if (CastSrcTy->isInteger())
9745 opcode = Instruction::IntToPtr;
9746 } else if (isa<IntegerType>(CastDstTy)) {
9747 if (isa<PointerType>(SIOp0->getType()))
9748 opcode = Instruction::PtrToInt;
9750 if (Constant *C = dyn_cast<Constant>(SIOp0))
9751 NewCast = ConstantExpr::getCast(opcode, C, CastDstTy);
9753 NewCast = IC.InsertNewInstBefore(
9754 CastInst::create(opcode, SIOp0, CastDstTy, SIOp0->getName()+".c"),
9756 return new StoreInst(NewCast, CastOp);
9763 Instruction *InstCombiner::visitStoreInst(StoreInst &SI) {
9764 Value *Val = SI.getOperand(0);
9765 Value *Ptr = SI.getOperand(1);
9767 if (isa<UndefValue>(Ptr)) { // store X, undef -> noop (even if volatile)
9768 EraseInstFromFunction(SI);
9773 // If the RHS is an alloca with a single use, zapify the store, making the
9775 if (Ptr->hasOneUse()) {
9776 if (isa<AllocaInst>(Ptr)) {
9777 EraseInstFromFunction(SI);
9782 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr))
9783 if (isa<AllocaInst>(GEP->getOperand(0)) &&
9784 GEP->getOperand(0)->hasOneUse()) {
9785 EraseInstFromFunction(SI);
9791 // Attempt to improve the alignment.
9792 unsigned KnownAlign = GetOrEnforceKnownAlignment(Ptr, TD);
9793 if (KnownAlign > SI.getAlignment())
9794 SI.setAlignment(KnownAlign);
9796 // Do really simple DSE, to catch cases where there are several consequtive
9797 // stores to the same location, separated by a few arithmetic operations. This
9798 // situation often occurs with bitfield accesses.
9799 BasicBlock::iterator BBI = &SI;
9800 for (unsigned ScanInsts = 6; BBI != SI.getParent()->begin() && ScanInsts;
9804 if (StoreInst *PrevSI = dyn_cast<StoreInst>(BBI)) {
9805 // Prev store isn't volatile, and stores to the same location?
9806 if (!PrevSI->isVolatile() && PrevSI->getOperand(1) == SI.getOperand(1)) {
9809 EraseInstFromFunction(*PrevSI);
9815 // If this is a load, we have to stop. However, if the loaded value is from
9816 // the pointer we're loading and is producing the pointer we're storing,
9817 // then *this* store is dead (X = load P; store X -> P).
9818 if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) {
9819 if (LI == Val && LI->getOperand(0) == Ptr && !SI.isVolatile()) {
9820 EraseInstFromFunction(SI);
9824 // Otherwise, this is a load from some other location. Stores before it
9829 // Don't skip over loads or things that can modify memory.
9830 if (BBI->mayWriteToMemory())
9835 if (SI.isVolatile()) return 0; // Don't hack volatile stores.
9837 // store X, null -> turns into 'unreachable' in SimplifyCFG
9838 if (isa<ConstantPointerNull>(Ptr)) {
9839 if (!isa<UndefValue>(Val)) {
9840 SI.setOperand(0, UndefValue::get(Val->getType()));
9841 if (Instruction *U = dyn_cast<Instruction>(Val))
9842 AddToWorkList(U); // Dropped a use.
9845 return 0; // Do not modify these!
9848 // store undef, Ptr -> noop
9849 if (isa<UndefValue>(Val)) {
9850 EraseInstFromFunction(SI);
9855 // If the pointer destination is a cast, see if we can fold the cast into the
9857 if (isa<CastInst>(Ptr))
9858 if (Instruction *Res = InstCombineStoreToCast(*this, SI))
9860 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ptr))
9862 if (Instruction *Res = InstCombineStoreToCast(*this, SI))
9866 // If this store is the last instruction in the basic block, and if the block
9867 // ends with an unconditional branch, try to move it to the successor block.
9869 if (BranchInst *BI = dyn_cast<BranchInst>(BBI))
9870 if (BI->isUnconditional())
9871 if (SimplifyStoreAtEndOfBlock(SI))
9872 return 0; // xform done!
9877 /// SimplifyStoreAtEndOfBlock - Turn things like:
9878 /// if () { *P = v1; } else { *P = v2 }
9879 /// into a phi node with a store in the successor.
9881 /// Simplify things like:
9882 /// *P = v1; if () { *P = v2; }
9883 /// into a phi node with a store in the successor.
9885 bool InstCombiner::SimplifyStoreAtEndOfBlock(StoreInst &SI) {
9886 BasicBlock *StoreBB = SI.getParent();
9888 // Check to see if the successor block has exactly two incoming edges. If
9889 // so, see if the other predecessor contains a store to the same location.
9890 // if so, insert a PHI node (if needed) and move the stores down.
9891 BasicBlock *DestBB = StoreBB->getTerminator()->getSuccessor(0);
9893 // Determine whether Dest has exactly two predecessors and, if so, compute
9894 // the other predecessor.
9895 pred_iterator PI = pred_begin(DestBB);
9896 BasicBlock *OtherBB = 0;
9900 if (PI == pred_end(DestBB))
9903 if (*PI != StoreBB) {
9908 if (++PI != pred_end(DestBB))
9912 // Verify that the other block ends in a branch and is not otherwise empty.
9913 BasicBlock::iterator BBI = OtherBB->getTerminator();
9914 BranchInst *OtherBr = dyn_cast<BranchInst>(BBI);
9915 if (!OtherBr || BBI == OtherBB->begin())
9918 // If the other block ends in an unconditional branch, check for the 'if then
9919 // else' case. there is an instruction before the branch.
9920 StoreInst *OtherStore = 0;
9921 if (OtherBr->isUnconditional()) {
9922 // If this isn't a store, or isn't a store to the same location, bail out.
9924 OtherStore = dyn_cast<StoreInst>(BBI);
9925 if (!OtherStore || OtherStore->getOperand(1) != SI.getOperand(1))
9928 // Otherwise, the other block ended with a conditional branch. If one of the
9929 // destinations is StoreBB, then we have the if/then case.
9930 if (OtherBr->getSuccessor(0) != StoreBB &&
9931 OtherBr->getSuccessor(1) != StoreBB)
9934 // Okay, we know that OtherBr now goes to Dest and StoreBB, so this is an
9935 // if/then triangle. See if there is a store to the same ptr as SI that
9936 // lives in OtherBB.
9938 // Check to see if we find the matching store.
9939 if ((OtherStore = dyn_cast<StoreInst>(BBI))) {
9940 if (OtherStore->getOperand(1) != SI.getOperand(1))
9944 // If we find something that may be using the stored value, or if we run
9945 // out of instructions, we can't do the xform.
9946 if (isa<LoadInst>(BBI) || BBI->mayWriteToMemory() ||
9947 BBI == OtherBB->begin())
9951 // In order to eliminate the store in OtherBr, we have to
9952 // make sure nothing reads the stored value in StoreBB.
9953 for (BasicBlock::iterator I = StoreBB->begin(); &*I != &SI; ++I) {
9954 // FIXME: This should really be AA driven.
9955 if (isa<LoadInst>(I) || I->mayWriteToMemory())
9960 // Insert a PHI node now if we need it.
9961 Value *MergedVal = OtherStore->getOperand(0);
9962 if (MergedVal != SI.getOperand(0)) {
9963 PHINode *PN = new PHINode(MergedVal->getType(), "storemerge");
9964 PN->reserveOperandSpace(2);
9965 PN->addIncoming(SI.getOperand(0), SI.getParent());
9966 PN->addIncoming(OtherStore->getOperand(0), OtherBB);
9967 MergedVal = InsertNewInstBefore(PN, DestBB->front());
9970 // Advance to a place where it is safe to insert the new store and
9972 BBI = DestBB->begin();
9973 while (isa<PHINode>(BBI)) ++BBI;
9974 InsertNewInstBefore(new StoreInst(MergedVal, SI.getOperand(1),
9975 OtherStore->isVolatile()), *BBI);
9977 // Nuke the old stores.
9978 EraseInstFromFunction(SI);
9979 EraseInstFromFunction(*OtherStore);
9985 Instruction *InstCombiner::visitBranchInst(BranchInst &BI) {
9986 // Change br (not X), label True, label False to: br X, label False, True
9988 BasicBlock *TrueDest;
9989 BasicBlock *FalseDest;
9990 if (match(&BI, m_Br(m_Not(m_Value(X)), TrueDest, FalseDest)) &&
9991 !isa<Constant>(X)) {
9992 // Swap Destinations and condition...
9994 BI.setSuccessor(0, FalseDest);
9995 BI.setSuccessor(1, TrueDest);
9999 // Cannonicalize fcmp_one -> fcmp_oeq
10000 FCmpInst::Predicate FPred; Value *Y;
10001 if (match(&BI, m_Br(m_FCmp(FPred, m_Value(X), m_Value(Y)),
10002 TrueDest, FalseDest)))
10003 if ((FPred == FCmpInst::FCMP_ONE || FPred == FCmpInst::FCMP_OLE ||
10004 FPred == FCmpInst::FCMP_OGE) && BI.getCondition()->hasOneUse()) {
10005 FCmpInst *I = cast<FCmpInst>(BI.getCondition());
10006 FCmpInst::Predicate NewPred = FCmpInst::getInversePredicate(FPred);
10007 Instruction *NewSCC = new FCmpInst(NewPred, X, Y, "", I);
10008 NewSCC->takeName(I);
10009 // Swap Destinations and condition...
10010 BI.setCondition(NewSCC);
10011 BI.setSuccessor(0, FalseDest);
10012 BI.setSuccessor(1, TrueDest);
10013 RemoveFromWorkList(I);
10014 I->eraseFromParent();
10015 AddToWorkList(NewSCC);
10019 // Cannonicalize icmp_ne -> icmp_eq
10020 ICmpInst::Predicate IPred;
10021 if (match(&BI, m_Br(m_ICmp(IPred, m_Value(X), m_Value(Y)),
10022 TrueDest, FalseDest)))
10023 if ((IPred == ICmpInst::ICMP_NE || IPred == ICmpInst::ICMP_ULE ||
10024 IPred == ICmpInst::ICMP_SLE || IPred == ICmpInst::ICMP_UGE ||
10025 IPred == ICmpInst::ICMP_SGE) && BI.getCondition()->hasOneUse()) {
10026 ICmpInst *I = cast<ICmpInst>(BI.getCondition());
10027 ICmpInst::Predicate NewPred = ICmpInst::getInversePredicate(IPred);
10028 Instruction *NewSCC = new ICmpInst(NewPred, X, Y, "", I);
10029 NewSCC->takeName(I);
10030 // Swap Destinations and condition...
10031 BI.setCondition(NewSCC);
10032 BI.setSuccessor(0, FalseDest);
10033 BI.setSuccessor(1, TrueDest);
10034 RemoveFromWorkList(I);
10035 I->eraseFromParent();;
10036 AddToWorkList(NewSCC);
10043 Instruction *InstCombiner::visitSwitchInst(SwitchInst &SI) {
10044 Value *Cond = SI.getCondition();
10045 if (Instruction *I = dyn_cast<Instruction>(Cond)) {
10046 if (I->getOpcode() == Instruction::Add)
10047 if (ConstantInt *AddRHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
10048 // change 'switch (X+4) case 1:' into 'switch (X) case -3'
10049 for (unsigned i = 2, e = SI.getNumOperands(); i != e; i += 2)
10050 SI.setOperand(i,ConstantExpr::getSub(cast<Constant>(SI.getOperand(i)),
10052 SI.setOperand(0, I->getOperand(0));
10060 /// CheapToScalarize - Return true if the value is cheaper to scalarize than it
10061 /// is to leave as a vector operation.
10062 static bool CheapToScalarize(Value *V, bool isConstant) {
10063 if (isa<ConstantAggregateZero>(V))
10065 if (ConstantVector *C = dyn_cast<ConstantVector>(V)) {
10066 if (isConstant) return true;
10067 // If all elts are the same, we can extract.
10068 Constant *Op0 = C->getOperand(0);
10069 for (unsigned i = 1; i < C->getNumOperands(); ++i)
10070 if (C->getOperand(i) != Op0)
10074 Instruction *I = dyn_cast<Instruction>(V);
10075 if (!I) return false;
10077 // Insert element gets simplified to the inserted element or is deleted if
10078 // this is constant idx extract element and its a constant idx insertelt.
10079 if (I->getOpcode() == Instruction::InsertElement && isConstant &&
10080 isa<ConstantInt>(I->getOperand(2)))
10082 if (I->getOpcode() == Instruction::Load && I->hasOneUse())
10084 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I))
10085 if (BO->hasOneUse() &&
10086 (CheapToScalarize(BO->getOperand(0), isConstant) ||
10087 CheapToScalarize(BO->getOperand(1), isConstant)))
10089 if (CmpInst *CI = dyn_cast<CmpInst>(I))
10090 if (CI->hasOneUse() &&
10091 (CheapToScalarize(CI->getOperand(0), isConstant) ||
10092 CheapToScalarize(CI->getOperand(1), isConstant)))
10098 /// Read and decode a shufflevector mask.
10100 /// It turns undef elements into values that are larger than the number of
10101 /// elements in the input.
10102 static std::vector<unsigned> getShuffleMask(const ShuffleVectorInst *SVI) {
10103 unsigned NElts = SVI->getType()->getNumElements();
10104 if (isa<ConstantAggregateZero>(SVI->getOperand(2)))
10105 return std::vector<unsigned>(NElts, 0);
10106 if (isa<UndefValue>(SVI->getOperand(2)))
10107 return std::vector<unsigned>(NElts, 2*NElts);
10109 std::vector<unsigned> Result;
10110 const ConstantVector *CP = cast<ConstantVector>(SVI->getOperand(2));
10111 for (unsigned i = 0, e = CP->getNumOperands(); i != e; ++i)
10112 if (isa<UndefValue>(CP->getOperand(i)))
10113 Result.push_back(NElts*2); // undef -> 8
10115 Result.push_back(cast<ConstantInt>(CP->getOperand(i))->getZExtValue());
10119 /// FindScalarElement - Given a vector and an element number, see if the scalar
10120 /// value is already around as a register, for example if it were inserted then
10121 /// extracted from the vector.
10122 static Value *FindScalarElement(Value *V, unsigned EltNo) {
10123 assert(isa<VectorType>(V->getType()) && "Not looking at a vector?");
10124 const VectorType *PTy = cast<VectorType>(V->getType());
10125 unsigned Width = PTy->getNumElements();
10126 if (EltNo >= Width) // Out of range access.
10127 return UndefValue::get(PTy->getElementType());
10129 if (isa<UndefValue>(V))
10130 return UndefValue::get(PTy->getElementType());
10131 else if (isa<ConstantAggregateZero>(V))
10132 return Constant::getNullValue(PTy->getElementType());
10133 else if (ConstantVector *CP = dyn_cast<ConstantVector>(V))
10134 return CP->getOperand(EltNo);
10135 else if (InsertElementInst *III = dyn_cast<InsertElementInst>(V)) {
10136 // If this is an insert to a variable element, we don't know what it is.
10137 if (!isa<ConstantInt>(III->getOperand(2)))
10139 unsigned IIElt = cast<ConstantInt>(III->getOperand(2))->getZExtValue();
10141 // If this is an insert to the element we are looking for, return the
10143 if (EltNo == IIElt)
10144 return III->getOperand(1);
10146 // Otherwise, the insertelement doesn't modify the value, recurse on its
10148 return FindScalarElement(III->getOperand(0), EltNo);
10149 } else if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(V)) {
10150 unsigned InEl = getShuffleMask(SVI)[EltNo];
10152 return FindScalarElement(SVI->getOperand(0), InEl);
10153 else if (InEl < Width*2)
10154 return FindScalarElement(SVI->getOperand(1), InEl - Width);
10156 return UndefValue::get(PTy->getElementType());
10159 // Otherwise, we don't know.
10163 Instruction *InstCombiner::visitExtractElementInst(ExtractElementInst &EI) {
10165 // If vector val is undef, replace extract with scalar undef.
10166 if (isa<UndefValue>(EI.getOperand(0)))
10167 return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType()));
10169 // If vector val is constant 0, replace extract with scalar 0.
10170 if (isa<ConstantAggregateZero>(EI.getOperand(0)))
10171 return ReplaceInstUsesWith(EI, Constant::getNullValue(EI.getType()));
10173 if (ConstantVector *C = dyn_cast<ConstantVector>(EI.getOperand(0))) {
10174 // If vector val is constant with uniform operands, replace EI
10175 // with that operand
10176 Constant *op0 = C->getOperand(0);
10177 for (unsigned i = 1; i < C->getNumOperands(); ++i)
10178 if (C->getOperand(i) != op0) {
10183 return ReplaceInstUsesWith(EI, op0);
10186 // If extracting a specified index from the vector, see if we can recursively
10187 // find a previously computed scalar that was inserted into the vector.
10188 if (ConstantInt *IdxC = dyn_cast<ConstantInt>(EI.getOperand(1))) {
10189 unsigned IndexVal = IdxC->getZExtValue();
10190 unsigned VectorWidth =
10191 cast<VectorType>(EI.getOperand(0)->getType())->getNumElements();
10193 // If this is extracting an invalid index, turn this into undef, to avoid
10194 // crashing the code below.
10195 if (IndexVal >= VectorWidth)
10196 return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType()));
10198 // This instruction only demands the single element from the input vector.
10199 // If the input vector has a single use, simplify it based on this use
10201 if (EI.getOperand(0)->hasOneUse() && VectorWidth != 1) {
10202 uint64_t UndefElts;
10203 if (Value *V = SimplifyDemandedVectorElts(EI.getOperand(0),
10206 EI.setOperand(0, V);
10211 if (Value *Elt = FindScalarElement(EI.getOperand(0), IndexVal))
10212 return ReplaceInstUsesWith(EI, Elt);
10214 // If the this extractelement is directly using a bitcast from a vector of
10215 // the same number of elements, see if we can find the source element from
10216 // it. In this case, we will end up needing to bitcast the scalars.
10217 if (BitCastInst *BCI = dyn_cast<BitCastInst>(EI.getOperand(0))) {
10218 if (const VectorType *VT =
10219 dyn_cast<VectorType>(BCI->getOperand(0)->getType()))
10220 if (VT->getNumElements() == VectorWidth)
10221 if (Value *Elt = FindScalarElement(BCI->getOperand(0), IndexVal))
10222 return new BitCastInst(Elt, EI.getType());
10226 if (Instruction *I = dyn_cast<Instruction>(EI.getOperand(0))) {
10227 if (I->hasOneUse()) {
10228 // Push extractelement into predecessor operation if legal and
10229 // profitable to do so
10230 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I)) {
10231 bool isConstantElt = isa<ConstantInt>(EI.getOperand(1));
10232 if (CheapToScalarize(BO, isConstantElt)) {
10233 ExtractElementInst *newEI0 =
10234 new ExtractElementInst(BO->getOperand(0), EI.getOperand(1),
10235 EI.getName()+".lhs");
10236 ExtractElementInst *newEI1 =
10237 new ExtractElementInst(BO->getOperand(1), EI.getOperand(1),
10238 EI.getName()+".rhs");
10239 InsertNewInstBefore(newEI0, EI);
10240 InsertNewInstBefore(newEI1, EI);
10241 return BinaryOperator::create(BO->getOpcode(), newEI0, newEI1);
10243 } else if (isa<LoadInst>(I)) {
10245 cast<PointerType>(I->getOperand(0)->getType())->getAddressSpace();
10246 Value *Ptr = InsertBitCastBefore(I->getOperand(0),
10247 PointerType::get(EI.getType(), AS),EI);
10248 GetElementPtrInst *GEP =
10249 new GetElementPtrInst(Ptr, EI.getOperand(1), I->getName() + ".gep");
10250 InsertNewInstBefore(GEP, EI);
10251 return new LoadInst(GEP);
10254 if (InsertElementInst *IE = dyn_cast<InsertElementInst>(I)) {
10255 // Extracting the inserted element?
10256 if (IE->getOperand(2) == EI.getOperand(1))
10257 return ReplaceInstUsesWith(EI, IE->getOperand(1));
10258 // If the inserted and extracted elements are constants, they must not
10259 // be the same value, extract from the pre-inserted value instead.
10260 if (isa<Constant>(IE->getOperand(2)) &&
10261 isa<Constant>(EI.getOperand(1))) {
10262 AddUsesToWorkList(EI);
10263 EI.setOperand(0, IE->getOperand(0));
10266 } else if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(I)) {
10267 // If this is extracting an element from a shufflevector, figure out where
10268 // it came from and extract from the appropriate input element instead.
10269 if (ConstantInt *Elt = dyn_cast<ConstantInt>(EI.getOperand(1))) {
10270 unsigned SrcIdx = getShuffleMask(SVI)[Elt->getZExtValue()];
10272 if (SrcIdx < SVI->getType()->getNumElements())
10273 Src = SVI->getOperand(0);
10274 else if (SrcIdx < SVI->getType()->getNumElements()*2) {
10275 SrcIdx -= SVI->getType()->getNumElements();
10276 Src = SVI->getOperand(1);
10278 return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType()));
10280 return new ExtractElementInst(Src, SrcIdx);
10287 /// CollectSingleShuffleElements - If V is a shuffle of values that ONLY returns
10288 /// elements from either LHS or RHS, return the shuffle mask and true.
10289 /// Otherwise, return false.
10290 static bool CollectSingleShuffleElements(Value *V, Value *LHS, Value *RHS,
10291 std::vector<Constant*> &Mask) {
10292 assert(V->getType() == LHS->getType() && V->getType() == RHS->getType() &&
10293 "Invalid CollectSingleShuffleElements");
10294 unsigned NumElts = cast<VectorType>(V->getType())->getNumElements();
10296 if (isa<UndefValue>(V)) {
10297 Mask.assign(NumElts, UndefValue::get(Type::Int32Ty));
10299 } else if (V == LHS) {
10300 for (unsigned i = 0; i != NumElts; ++i)
10301 Mask.push_back(ConstantInt::get(Type::Int32Ty, i));
10303 } else if (V == RHS) {
10304 for (unsigned i = 0; i != NumElts; ++i)
10305 Mask.push_back(ConstantInt::get(Type::Int32Ty, i+NumElts));
10307 } else if (InsertElementInst *IEI = dyn_cast<InsertElementInst>(V)) {
10308 // If this is an insert of an extract from some other vector, include it.
10309 Value *VecOp = IEI->getOperand(0);
10310 Value *ScalarOp = IEI->getOperand(1);
10311 Value *IdxOp = IEI->getOperand(2);
10313 if (!isa<ConstantInt>(IdxOp))
10315 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
10317 if (isa<UndefValue>(ScalarOp)) { // inserting undef into vector.
10318 // Okay, we can handle this if the vector we are insertinting into is
10319 // transitively ok.
10320 if (CollectSingleShuffleElements(VecOp, LHS, RHS, Mask)) {
10321 // If so, update the mask to reflect the inserted undef.
10322 Mask[InsertedIdx] = UndefValue::get(Type::Int32Ty);
10325 } else if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)){
10326 if (isa<ConstantInt>(EI->getOperand(1)) &&
10327 EI->getOperand(0)->getType() == V->getType()) {
10328 unsigned ExtractedIdx =
10329 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
10331 // This must be extracting from either LHS or RHS.
10332 if (EI->getOperand(0) == LHS || EI->getOperand(0) == RHS) {
10333 // Okay, we can handle this if the vector we are insertinting into is
10334 // transitively ok.
10335 if (CollectSingleShuffleElements(VecOp, LHS, RHS, Mask)) {
10336 // If so, update the mask to reflect the inserted value.
10337 if (EI->getOperand(0) == LHS) {
10338 Mask[InsertedIdx & (NumElts-1)] =
10339 ConstantInt::get(Type::Int32Ty, ExtractedIdx);
10341 assert(EI->getOperand(0) == RHS);
10342 Mask[InsertedIdx & (NumElts-1)] =
10343 ConstantInt::get(Type::Int32Ty, ExtractedIdx+NumElts);
10352 // TODO: Handle shufflevector here!
10357 /// CollectShuffleElements - We are building a shuffle of V, using RHS as the
10358 /// RHS of the shuffle instruction, if it is not null. Return a shuffle mask
10359 /// that computes V and the LHS value of the shuffle.
10360 static Value *CollectShuffleElements(Value *V, std::vector<Constant*> &Mask,
10362 assert(isa<VectorType>(V->getType()) &&
10363 (RHS == 0 || V->getType() == RHS->getType()) &&
10364 "Invalid shuffle!");
10365 unsigned NumElts = cast<VectorType>(V->getType())->getNumElements();
10367 if (isa<UndefValue>(V)) {
10368 Mask.assign(NumElts, UndefValue::get(Type::Int32Ty));
10370 } else if (isa<ConstantAggregateZero>(V)) {
10371 Mask.assign(NumElts, ConstantInt::get(Type::Int32Ty, 0));
10373 } else if (InsertElementInst *IEI = dyn_cast<InsertElementInst>(V)) {
10374 // If this is an insert of an extract from some other vector, include it.
10375 Value *VecOp = IEI->getOperand(0);
10376 Value *ScalarOp = IEI->getOperand(1);
10377 Value *IdxOp = IEI->getOperand(2);
10379 if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)) {
10380 if (isa<ConstantInt>(EI->getOperand(1)) && isa<ConstantInt>(IdxOp) &&
10381 EI->getOperand(0)->getType() == V->getType()) {
10382 unsigned ExtractedIdx =
10383 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
10384 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
10386 // Either the extracted from or inserted into vector must be RHSVec,
10387 // otherwise we'd end up with a shuffle of three inputs.
10388 if (EI->getOperand(0) == RHS || RHS == 0) {
10389 RHS = EI->getOperand(0);
10390 Value *V = CollectShuffleElements(VecOp, Mask, RHS);
10391 Mask[InsertedIdx & (NumElts-1)] =
10392 ConstantInt::get(Type::Int32Ty, NumElts+ExtractedIdx);
10396 if (VecOp == RHS) {
10397 Value *V = CollectShuffleElements(EI->getOperand(0), Mask, RHS);
10398 // Everything but the extracted element is replaced with the RHS.
10399 for (unsigned i = 0; i != NumElts; ++i) {
10400 if (i != InsertedIdx)
10401 Mask[i] = ConstantInt::get(Type::Int32Ty, NumElts+i);
10406 // If this insertelement is a chain that comes from exactly these two
10407 // vectors, return the vector and the effective shuffle.
10408 if (CollectSingleShuffleElements(IEI, EI->getOperand(0), RHS, Mask))
10409 return EI->getOperand(0);
10414 // TODO: Handle shufflevector here!
10416 // Otherwise, can't do anything fancy. Return an identity vector.
10417 for (unsigned i = 0; i != NumElts; ++i)
10418 Mask.push_back(ConstantInt::get(Type::Int32Ty, i));
10422 Instruction *InstCombiner::visitInsertElementInst(InsertElementInst &IE) {
10423 Value *VecOp = IE.getOperand(0);
10424 Value *ScalarOp = IE.getOperand(1);
10425 Value *IdxOp = IE.getOperand(2);
10427 // Inserting an undef or into an undefined place, remove this.
10428 if (isa<UndefValue>(ScalarOp) || isa<UndefValue>(IdxOp))
10429 ReplaceInstUsesWith(IE, VecOp);
10431 // If the inserted element was extracted from some other vector, and if the
10432 // indexes are constant, try to turn this into a shufflevector operation.
10433 if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)) {
10434 if (isa<ConstantInt>(EI->getOperand(1)) && isa<ConstantInt>(IdxOp) &&
10435 EI->getOperand(0)->getType() == IE.getType()) {
10436 unsigned NumVectorElts = IE.getType()->getNumElements();
10437 unsigned ExtractedIdx =
10438 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
10439 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
10441 if (ExtractedIdx >= NumVectorElts) // Out of range extract.
10442 return ReplaceInstUsesWith(IE, VecOp);
10444 if (InsertedIdx >= NumVectorElts) // Out of range insert.
10445 return ReplaceInstUsesWith(IE, UndefValue::get(IE.getType()));
10447 // If we are extracting a value from a vector, then inserting it right
10448 // back into the same place, just use the input vector.
10449 if (EI->getOperand(0) == VecOp && ExtractedIdx == InsertedIdx)
10450 return ReplaceInstUsesWith(IE, VecOp);
10452 // We could theoretically do this for ANY input. However, doing so could
10453 // turn chains of insertelement instructions into a chain of shufflevector
10454 // instructions, and right now we do not merge shufflevectors. As such,
10455 // only do this in a situation where it is clear that there is benefit.
10456 if (isa<UndefValue>(VecOp) || isa<ConstantAggregateZero>(VecOp)) {
10457 // Turn this into shuffle(EIOp0, VecOp, Mask). The result has all of
10458 // the values of VecOp, except then one read from EIOp0.
10459 // Build a new shuffle mask.
10460 std::vector<Constant*> Mask;
10461 if (isa<UndefValue>(VecOp))
10462 Mask.assign(NumVectorElts, UndefValue::get(Type::Int32Ty));
10464 assert(isa<ConstantAggregateZero>(VecOp) && "Unknown thing");
10465 Mask.assign(NumVectorElts, ConstantInt::get(Type::Int32Ty,
10468 Mask[InsertedIdx] = ConstantInt::get(Type::Int32Ty, ExtractedIdx);
10469 return new ShuffleVectorInst(EI->getOperand(0), VecOp,
10470 ConstantVector::get(Mask));
10473 // If this insertelement isn't used by some other insertelement, turn it
10474 // (and any insertelements it points to), into one big shuffle.
10475 if (!IE.hasOneUse() || !isa<InsertElementInst>(IE.use_back())) {
10476 std::vector<Constant*> Mask;
10478 Value *LHS = CollectShuffleElements(&IE, Mask, RHS);
10479 if (RHS == 0) RHS = UndefValue::get(LHS->getType());
10480 // We now have a shuffle of LHS, RHS, Mask.
10481 return new ShuffleVectorInst(LHS, RHS, ConstantVector::get(Mask));
10490 Instruction *InstCombiner::visitShuffleVectorInst(ShuffleVectorInst &SVI) {
10491 Value *LHS = SVI.getOperand(0);
10492 Value *RHS = SVI.getOperand(1);
10493 std::vector<unsigned> Mask = getShuffleMask(&SVI);
10495 bool MadeChange = false;
10497 // Undefined shuffle mask -> undefined value.
10498 if (isa<UndefValue>(SVI.getOperand(2)))
10499 return ReplaceInstUsesWith(SVI, UndefValue::get(SVI.getType()));
10501 // If we have shuffle(x, undef, mask) and any elements of mask refer to
10502 // the undef, change them to undefs.
10503 if (isa<UndefValue>(SVI.getOperand(1))) {
10504 // Scan to see if there are any references to the RHS. If so, replace them
10505 // with undef element refs and set MadeChange to true.
10506 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
10507 if (Mask[i] >= e && Mask[i] != 2*e) {
10514 // Remap any references to RHS to use LHS.
10515 std::vector<Constant*> Elts;
10516 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
10517 if (Mask[i] == 2*e)
10518 Elts.push_back(UndefValue::get(Type::Int32Ty));
10520 Elts.push_back(ConstantInt::get(Type::Int32Ty, Mask[i]));
10522 SVI.setOperand(2, ConstantVector::get(Elts));
10526 // Canonicalize shuffle(x ,x,mask) -> shuffle(x, undef,mask')
10527 // Canonicalize shuffle(undef,x,mask) -> shuffle(x, undef,mask').
10528 if (LHS == RHS || isa<UndefValue>(LHS)) {
10529 if (isa<UndefValue>(LHS) && LHS == RHS) {
10530 // shuffle(undef,undef,mask) -> undef.
10531 return ReplaceInstUsesWith(SVI, LHS);
10534 // Remap any references to RHS to use LHS.
10535 std::vector<Constant*> Elts;
10536 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
10537 if (Mask[i] >= 2*e)
10538 Elts.push_back(UndefValue::get(Type::Int32Ty));
10540 if ((Mask[i] >= e && isa<UndefValue>(RHS)) ||
10541 (Mask[i] < e && isa<UndefValue>(LHS)))
10542 Mask[i] = 2*e; // Turn into undef.
10544 Mask[i] &= (e-1); // Force to LHS.
10545 Elts.push_back(ConstantInt::get(Type::Int32Ty, Mask[i]));
10548 SVI.setOperand(0, SVI.getOperand(1));
10549 SVI.setOperand(1, UndefValue::get(RHS->getType()));
10550 SVI.setOperand(2, ConstantVector::get(Elts));
10551 LHS = SVI.getOperand(0);
10552 RHS = SVI.getOperand(1);
10556 // Analyze the shuffle, are the LHS or RHS and identity shuffles?
10557 bool isLHSID = true, isRHSID = true;
10559 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
10560 if (Mask[i] >= e*2) continue; // Ignore undef values.
10561 // Is this an identity shuffle of the LHS value?
10562 isLHSID &= (Mask[i] == i);
10564 // Is this an identity shuffle of the RHS value?
10565 isRHSID &= (Mask[i]-e == i);
10568 // Eliminate identity shuffles.
10569 if (isLHSID) return ReplaceInstUsesWith(SVI, LHS);
10570 if (isRHSID) return ReplaceInstUsesWith(SVI, RHS);
10572 // If the LHS is a shufflevector itself, see if we can combine it with this
10573 // one without producing an unusual shuffle. Here we are really conservative:
10574 // we are absolutely afraid of producing a shuffle mask not in the input
10575 // program, because the code gen may not be smart enough to turn a merged
10576 // shuffle into two specific shuffles: it may produce worse code. As such,
10577 // we only merge two shuffles if the result is one of the two input shuffle
10578 // masks. In this case, merging the shuffles just removes one instruction,
10579 // which we know is safe. This is good for things like turning:
10580 // (splat(splat)) -> splat.
10581 if (ShuffleVectorInst *LHSSVI = dyn_cast<ShuffleVectorInst>(LHS)) {
10582 if (isa<UndefValue>(RHS)) {
10583 std::vector<unsigned> LHSMask = getShuffleMask(LHSSVI);
10585 std::vector<unsigned> NewMask;
10586 for (unsigned i = 0, e = Mask.size(); i != e; ++i)
10587 if (Mask[i] >= 2*e)
10588 NewMask.push_back(2*e);
10590 NewMask.push_back(LHSMask[Mask[i]]);
10592 // If the result mask is equal to the src shuffle or this shuffle mask, do
10593 // the replacement.
10594 if (NewMask == LHSMask || NewMask == Mask) {
10595 std::vector<Constant*> Elts;
10596 for (unsigned i = 0, e = NewMask.size(); i != e; ++i) {
10597 if (NewMask[i] >= e*2) {
10598 Elts.push_back(UndefValue::get(Type::Int32Ty));
10600 Elts.push_back(ConstantInt::get(Type::Int32Ty, NewMask[i]));
10603 return new ShuffleVectorInst(LHSSVI->getOperand(0),
10604 LHSSVI->getOperand(1),
10605 ConstantVector::get(Elts));
10610 return MadeChange ? &SVI : 0;
10616 /// TryToSinkInstruction - Try to move the specified instruction from its
10617 /// current block into the beginning of DestBlock, which can only happen if it's
10618 /// safe to move the instruction past all of the instructions between it and the
10619 /// end of its block.
10620 static bool TryToSinkInstruction(Instruction *I, BasicBlock *DestBlock) {
10621 assert(I->hasOneUse() && "Invariants didn't hold!");
10623 // Cannot move control-flow-involving, volatile loads, vaarg, etc.
10624 if (isa<PHINode>(I) || I->mayWriteToMemory()) return false;
10626 // Do not sink alloca instructions out of the entry block.
10627 if (isa<AllocaInst>(I) && I->getParent() ==
10628 &DestBlock->getParent()->getEntryBlock())
10631 // We can only sink load instructions if there is nothing between the load and
10632 // the end of block that could change the value.
10633 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
10634 for (BasicBlock::iterator Scan = LI, E = LI->getParent()->end();
10636 if (Scan->mayWriteToMemory())
10640 BasicBlock::iterator InsertPos = DestBlock->begin();
10641 while (isa<PHINode>(InsertPos)) ++InsertPos;
10643 I->moveBefore(InsertPos);
10649 /// AddReachableCodeToWorklist - Walk the function in depth-first order, adding
10650 /// all reachable code to the worklist.
10652 /// This has a couple of tricks to make the code faster and more powerful. In
10653 /// particular, we constant fold and DCE instructions as we go, to avoid adding
10654 /// them to the worklist (this significantly speeds up instcombine on code where
10655 /// many instructions are dead or constant). Additionally, if we find a branch
10656 /// whose condition is a known constant, we only visit the reachable successors.
10658 static void AddReachableCodeToWorklist(BasicBlock *BB,
10659 SmallPtrSet<BasicBlock*, 64> &Visited,
10661 const TargetData *TD) {
10662 std::vector<BasicBlock*> Worklist;
10663 Worklist.push_back(BB);
10665 while (!Worklist.empty()) {
10666 BB = Worklist.back();
10667 Worklist.pop_back();
10669 // We have now visited this block! If we've already been here, ignore it.
10670 if (!Visited.insert(BB)) continue;
10672 for (BasicBlock::iterator BBI = BB->begin(), E = BB->end(); BBI != E; ) {
10673 Instruction *Inst = BBI++;
10675 // DCE instruction if trivially dead.
10676 if (isInstructionTriviallyDead(Inst)) {
10678 DOUT << "IC: DCE: " << *Inst;
10679 Inst->eraseFromParent();
10683 // ConstantProp instruction if trivially constant.
10684 if (Constant *C = ConstantFoldInstruction(Inst, TD)) {
10685 DOUT << "IC: ConstFold to: " << *C << " from: " << *Inst;
10686 Inst->replaceAllUsesWith(C);
10688 Inst->eraseFromParent();
10692 IC.AddToWorkList(Inst);
10695 // Recursively visit successors. If this is a branch or switch on a
10696 // constant, only visit the reachable successor.
10697 TerminatorInst *TI = BB->getTerminator();
10698 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
10699 if (BI->isConditional() && isa<ConstantInt>(BI->getCondition())) {
10700 bool CondVal = cast<ConstantInt>(BI->getCondition())->getZExtValue();
10701 Worklist.push_back(BI->getSuccessor(!CondVal));
10704 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
10705 if (ConstantInt *Cond = dyn_cast<ConstantInt>(SI->getCondition())) {
10706 // See if this is an explicit destination.
10707 for (unsigned i = 1, e = SI->getNumSuccessors(); i != e; ++i)
10708 if (SI->getCaseValue(i) == Cond) {
10709 Worklist.push_back(SI->getSuccessor(i));
10713 // Otherwise it is the default destination.
10714 Worklist.push_back(SI->getSuccessor(0));
10719 for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i)
10720 Worklist.push_back(TI->getSuccessor(i));
10724 bool InstCombiner::DoOneIteration(Function &F, unsigned Iteration) {
10725 bool Changed = false;
10726 TD = &getAnalysis<TargetData>();
10728 DEBUG(DOUT << "\n\nINSTCOMBINE ITERATION #" << Iteration << " on "
10729 << F.getNameStr() << "\n");
10732 // Do a depth-first traversal of the function, populate the worklist with
10733 // the reachable instructions. Ignore blocks that are not reachable. Keep
10734 // track of which blocks we visit.
10735 SmallPtrSet<BasicBlock*, 64> Visited;
10736 AddReachableCodeToWorklist(F.begin(), Visited, *this, TD);
10738 // Do a quick scan over the function. If we find any blocks that are
10739 // unreachable, remove any instructions inside of them. This prevents
10740 // the instcombine code from having to deal with some bad special cases.
10741 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB)
10742 if (!Visited.count(BB)) {
10743 Instruction *Term = BB->getTerminator();
10744 while (Term != BB->begin()) { // Remove instrs bottom-up
10745 BasicBlock::iterator I = Term; --I;
10747 DOUT << "IC: DCE: " << *I;
10750 if (!I->use_empty())
10751 I->replaceAllUsesWith(UndefValue::get(I->getType()));
10752 I->eraseFromParent();
10757 while (!Worklist.empty()) {
10758 Instruction *I = RemoveOneFromWorkList();
10759 if (I == 0) continue; // skip null values.
10761 // Check to see if we can DCE the instruction.
10762 if (isInstructionTriviallyDead(I)) {
10763 // Add operands to the worklist.
10764 if (I->getNumOperands() < 4)
10765 AddUsesToWorkList(*I);
10768 DOUT << "IC: DCE: " << *I;
10770 I->eraseFromParent();
10771 RemoveFromWorkList(I);
10775 // Instruction isn't dead, see if we can constant propagate it.
10776 if (Constant *C = ConstantFoldInstruction(I, TD)) {
10777 DOUT << "IC: ConstFold to: " << *C << " from: " << *I;
10779 // Add operands to the worklist.
10780 AddUsesToWorkList(*I);
10781 ReplaceInstUsesWith(*I, C);
10784 I->eraseFromParent();
10785 RemoveFromWorkList(I);
10789 // See if we can trivially sink this instruction to a successor basic block.
10790 if (I->hasOneUse()) {
10791 BasicBlock *BB = I->getParent();
10792 BasicBlock *UserParent = cast<Instruction>(I->use_back())->getParent();
10793 if (UserParent != BB) {
10794 bool UserIsSuccessor = false;
10795 // See if the user is one of our successors.
10796 for (succ_iterator SI = succ_begin(BB), E = succ_end(BB); SI != E; ++SI)
10797 if (*SI == UserParent) {
10798 UserIsSuccessor = true;
10802 // If the user is one of our immediate successors, and if that successor
10803 // only has us as a predecessors (we'd have to split the critical edge
10804 // otherwise), we can keep going.
10805 if (UserIsSuccessor && !isa<PHINode>(I->use_back()) &&
10806 next(pred_begin(UserParent)) == pred_end(UserParent))
10807 // Okay, the CFG is simple enough, try to sink this instruction.
10808 Changed |= TryToSinkInstruction(I, UserParent);
10812 // Now that we have an instruction, try combining it to simplify it...
10816 DEBUG(std::ostringstream SS; I->print(SS); OrigI = SS.str(););
10817 if (Instruction *Result = visit(*I)) {
10819 // Should we replace the old instruction with a new one?
10821 DOUT << "IC: Old = " << *I
10822 << " New = " << *Result;
10824 // Everything uses the new instruction now.
10825 I->replaceAllUsesWith(Result);
10827 // Push the new instruction and any users onto the worklist.
10828 AddToWorkList(Result);
10829 AddUsersToWorkList(*Result);
10831 // Move the name to the new instruction first.
10832 Result->takeName(I);
10834 // Insert the new instruction into the basic block...
10835 BasicBlock *InstParent = I->getParent();
10836 BasicBlock::iterator InsertPos = I;
10838 if (!isa<PHINode>(Result)) // If combining a PHI, don't insert
10839 while (isa<PHINode>(InsertPos)) // middle of a block of PHIs.
10842 InstParent->getInstList().insert(InsertPos, Result);
10844 // Make sure that we reprocess all operands now that we reduced their
10846 AddUsesToWorkList(*I);
10848 // Instructions can end up on the worklist more than once. Make sure
10849 // we do not process an instruction that has been deleted.
10850 RemoveFromWorkList(I);
10852 // Erase the old instruction.
10853 InstParent->getInstList().erase(I);
10856 DOUT << "IC: Mod = " << OrigI
10857 << " New = " << *I;
10860 // If the instruction was modified, it's possible that it is now dead.
10861 // if so, remove it.
10862 if (isInstructionTriviallyDead(I)) {
10863 // Make sure we process all operands now that we are reducing their
10865 AddUsesToWorkList(*I);
10867 // Instructions may end up in the worklist more than once. Erase all
10868 // occurrences of this instruction.
10869 RemoveFromWorkList(I);
10870 I->eraseFromParent();
10873 AddUsersToWorkList(*I);
10880 assert(WorklistMap.empty() && "Worklist empty, but map not?");
10882 // Do an explicit clear, this shrinks the map if needed.
10883 WorklistMap.clear();
10888 bool InstCombiner::runOnFunction(Function &F) {
10889 MustPreserveLCSSA = mustPreserveAnalysisID(LCSSAID);
10891 bool EverMadeChange = false;
10893 // Iterate while there is work to do.
10894 unsigned Iteration = 0;
10895 while (DoOneIteration(F, Iteration++))
10896 EverMadeChange = true;
10897 return EverMadeChange;
10900 FunctionPass *llvm::createInstructionCombiningPass() {
10901 return new InstCombiner();